From e393515c7c94e0f127bcbac633052ee457c0527e Mon Sep 17 00:00:00 2001
From: Agent IX <agent@agent-ix.dev>
Date: Sat, 29 Aug 2026 02:50:47 -0700
Subject: [PATCH 1/2] fix: prevent workstation path publication

---
 .github/workflows/path-hygiene.yml            |  36 +++++
 AGENTS.md                                     |  11 ++
 .../install-pgvectorscale-release.log         |   8 +-
 .../feedback/2026-04-08-01-reviewer.md        |   3 +-
 crates/ecaz-cli/README.md                     |  15 +-
 crates/ecaz-cli/src/commands/dev/install.rs   |  16 +--
 ...re-phase11-paper-parity-production-gate.md |   2 +-
 plan/tasks/240-artifact-path-sanitization.md  |  42 ++++++
 ...87-candidate-batched-scoring-across-ams.md |   4 +-
 .../88-streaming-ann-result-iteration.md      |   6 +-
 ...9-turboquant-tqplus-cross-am-validation.md |   2 +-
 plan/tasks/README.md                          |   2 +
 .../request.md                                |   2 +-
 .../feedback/2026-04-09-01-reviewer.md        |  12 +-
 .../request.md                                |   3 +-
 .../request.md                                |   3 +-
 .../artifacts/manifest.md                     |   2 +-
 .../feedback/2026-07-30-02-reviewer.md        |   2 +-
 .../feedback/2026-07-30-02-reviewer.md        |   2 +-
 .../artifacts/paper-algorithm-citations.md    |   2 +-
 .../003-ab/feedback/2026-07-30-02-reviewer.md |   2 +-
 .../request.md                                |   8 +-
 .../feedback/2026-04-23-01-reviewer.md        |   2 +-
 .../feedback.md                               |   2 +-
 .../feedback/2026-05-09-01-reviewer.md        |   2 +-
 .../artifacts/paper-parity-seed.md            |   2 +-
 .../artifacts/paper-source-outline.md         |   2 +-
 .../artifacts/manifest.md                     |   2 +-
 .../artifacts/turbovec-tq-analysis.md         |  40 +++---
 .../artifacts/source-scoring-map.md           |   2 +-
 scripts/check_workstation_paths.py            | 136 ++++++++++++++++++
 scripts/tests/test_check_workstation_paths.py |  74 ++++++++++
 spec/adr/ADR-006-own-quantizer.md             |   3 +-
 spec/adr/ADR-024-fwht-transform-strategy.md   |   5 +-
 ...ymphonyqg-quantized-graph-access-method.md |   3 +-
 .../ADR-085-ec-distann-single-global-graph.md |   3 +-
 spec/reviews/evidence.md                      |   2 +-
 37 files changed, 379 insertions(+), 86 deletions(-)
 create mode 100644 .github/workflows/path-hygiene.yml
 create mode 100644 plan/tasks/240-artifact-path-sanitization.md
 create mode 100644 scripts/check_workstation_paths.py
 create mode 100644 scripts/tests/test_check_workstation_paths.py

diff --git a/.github/workflows/path-hygiene.yml b/.github/workflows/path-hygiene.yml
new file mode 100644
index 0000000000..db36376d0d
--- /dev/null
+++ b/.github/workflows/path-hygiene.yml
@@ -0,0 +1,36 @@
+name: Path hygiene
+
+on:
+  pull_request:
+  push:
+    branches: [main]
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+jobs:
+  added-lines:
+    runs-on: ubuntu-24.04
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+
+      - name: Reject newly added workstation paths
+        env:
+          EVENT_NAME: ${{ github.event_name }}
+          BASE_SHA: ${{ github.event.pull_request.base.sha }}
+          BEFORE_SHA: ${{ github.event.before }}
+        run: |
+          if [ "$EVENT_NAME" = "pull_request" ]; then
+            diff_range="${BASE_SHA}...HEAD"
+          elif [ -n "$BEFORE_SHA" ] && ! printf '%s' "$BEFORE_SHA" | grep -Eq '^0+$'; then
+            diff_range="${BEFORE_SHA}..HEAD"
+          elif git rev-parse HEAD^ >/dev/null 2>&1; then
+            diff_range="HEAD^..HEAD"
+          else
+            empty_tree="$(git hash-object -t tree /dev/null)"
+            diff_range="${empty_tree}..HEAD"
+          fi
+          python3 scripts/check_workstation_paths.py --diff-range "$diff_range"
diff --git a/AGENTS.md b/AGENTS.md
index d3f120aeb6..4d90dd5a5e 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -94,6 +94,17 @@ Packet directories inside a task bucket must sort in chronological order.
 - `request.md` should summarize the result and point at the packet-local
   artifact files.
 
+### Workstation Path Hygiene
+
+- New or changed tracked content must not introduce a machine-specific absolute
+  workstation home path. Use an environment variable, repository-relative path,
+  or durable public URL instead.
+- `scripts/check_workstation_paths.py` enforces this on added diff lines without
+  echoing the sensitive value into CI output.
+- Historical evidence that predates this control must not be bulk-rewritten merely
+  to make the legacy corpus look clean. Any redaction requires an explicit audit
+  trail and a separate history-remediation decision.
+
 ### Never Commit: Corpus Data, Operational Logs, and Polling Cruft
 
 A review packet is decision-grade evidence, not a capture of everything the run
diff --git a/benchmarks/11105-task29-release-latency-refresh/artifacts/install-pgvectorscale-release.log b/benchmarks/11105-task29-release-latency-refresh/artifacts/install-pgvectorscale-release.log
index 0a2d74006a..eacdbe86e3 100644
--- a/benchmarks/11105-task29-release-latency-refresh/artifacts/install-pgvectorscale-release.log
+++ b/benchmarks/11105-task29-release-latency-refresh/artifacts/install-pgvectorscale-release.log
@@ -118,7 +118,7 @@ Script started on 2026-04-30 16:13:15-07:00 [<not executed on terminal>]
    Compiling ptr_meta_derive v0.1.4
    Compiling bytecheck_derive v0.6.12
    Compiling rkyv_derive v0.7.46
-   Compiling pgvectorscale_derive v0.1.0 (/home/peter/dev_bak/pgvectorscale/pgvectorscale/pgvectorscale_derive)
+   Compiling pgvectorscale_derive v0.1.0 ([local-checkout]/pgvectorscale/pgvectorscale_derive)
    Compiling ptr_meta v0.1.4
    Compiling synstructure v0.13.2
    Compiling futures-executor v0.3.32
@@ -157,7 +157,7 @@ Script started on 2026-04-30 16:13:15-07:00 [<not executed on terminal>]
    Compiling pgrx-bindgen v0.16.1
    Compiling pgrx-pg-sys v0.16.1
    Compiling pgrx v0.16.1
-   Compiling vectorscale v0.9.0 (/home/peter/dev_bak/pgvectorscale/pgvectorscale)
+   Compiling vectorscale v0.9.0 ([local-checkout]/pgvectorscale)
     Finished `release` profile [optimized] target(s) in 54.98s
   Installing extension
      Copying control file to /home/peter/.pgrx/18.3/pgrx-install/share/postgresql/extension/vectorscale.control
@@ -279,7 +279,7 @@ Script started on 2026-04-30 16:13:15-07:00 [<not executed on terminal>]
    Compiling ptr_meta_derive v0.1.4
    Compiling bytecheck_derive v0.6.12
    Compiling rkyv_derive v0.7.46
-   Compiling pgvectorscale_derive v0.1.0 (/home/peter/dev_bak/pgvectorscale/pgvectorscale/pgvectorscale_derive)
+   Compiling pgvectorscale_derive v0.1.0 ([local-checkout]/pgvectorscale/pgvectorscale_derive)
    Compiling futures-executor v0.3.32
    Compiling ptr_meta v0.1.4
    Compiling regex v1.12.3
@@ -319,7 +319,7 @@ Script started on 2026-04-30 16:13:15-07:00 [<not executed on terminal>]
    Compiling pgrx-bindgen v0.16.1
    Compiling pgrx-pg-sys v0.16.1
    Compiling pgrx v0.16.1
-   Compiling vectorscale v0.9.0 (/home/peter/dev_bak/pgvectorscale/pgvectorscale)
+   Compiling vectorscale v0.9.0 ([local-checkout]/pgvectorscale)
     Finished `dev` profile [unoptimized + debuginfo] target(s) in 37.88s
      Writing SQL entities to /home/peter/.pgrx/18.3/pgrx-install/share/postgresql/extension/vectorscale--0.9.0.sql
      Copying extension schema upgrade file to /home/peter/.pgrx/18.3/pgrx-install/share/postgresql/extension/vectorscale--0.2.0--0.3.0.sql
diff --git a/benchmarks/200-a4-recall-gate-rerun/feedback/2026-04-08-01-reviewer.md b/benchmarks/200-a4-recall-gate-rerun/feedback/2026-04-08-01-reviewer.md
index f6d03916aa..a91ea3e810 100644
--- a/benchmarks/200-a4-recall-gate-rerun/feedback/2026-04-08-01-reviewer.md
+++ b/benchmarks/200-a4-recall-gate-rerun/feedback/2026-04-08-01-reviewer.md
@@ -72,7 +72,8 @@ Well-framed. Updated assessment after cross-referencing reviews 195-198 and 201:
 
 ## Additional Findings
 
-Verified against reference implementations in `~/dev_bak/hnsw-references/`:
+Verified against local reference checkouts of `hnsw`, `hnswlib-rs`,
+`instant-distance`, and `swarc`:
 
 1. **Layer indexing**: hnswlib-rs `get_neighborhood_id()` iterates `0..nb_layer` — layer 0 = index 0.
    Matches `pack_point_neighbor_slots`. No reversal bug.
diff --git a/crates/ecaz-cli/README.md b/crates/ecaz-cli/README.md
index d57a589c3b..4081136d22 100644
--- a/crates/ecaz-cli/README.md
+++ b/crates/ecaz-cli/README.md
@@ -47,14 +47,14 @@ cargo install --path crates/ecaz-cli
 That installs the binary under Cargo's bin directory, usually
 `$HOME/.cargo/bin/ecaz`. Interactive shells may also put that directory on
 `$PATH`; when running from an agent or other sandboxed tool session, prefer the
-absolute installed path (for example `/Users/peter/.cargo/bin/ecaz`) so one
+absolute installed path (for example `${HOME}/.cargo/bin/ecaz`) so one
 approval rule can cover the operator surface consistently.
 
 All commands accept `--database`, `--host`, `--port`, `--user`, `--password`,
 and `--log-file`; each also falls back to the matching libpq environment
 variable (`PGDATABASE`, `PGHOST`, `PGPORT`, `PGUSER`, `PGPASSWORD`) when the
 flag is omitted. `--host` may be either a TCP host name or a Unix socket
-directory such as `/home/peter/.pgrx`.
+directory supplied through `PGRX_HOME` or `PGHOST`.
 `--log-file` mirrors the CLI's stdout/stderr into a packet-local artifact
 file so review runs do not need shell `tee` wrappers. When `--log-file` is
 set, transient progress bars are suppressed so the artifact stays stable.
@@ -67,8 +67,8 @@ generation, load/list/inspect, benchmark, storage, scratch, and SQL checks
 through that binary:
 
 ```sh
-/Users/peter/.cargo/bin/ecaz dev sql --pg 18 --db postgres \
-  --socket-dir /Users/peter/.pgrx --raw \
+${HOME}/.cargo/bin/ecaz dev sql --pg 18 --db postgres \
+  --socket-dir "${PGRX_HOME}" --raw \
   --sql "select version()" \
   --log-output reviews/task-{id}/001-example/artifacts/pg18-status.log
 ```
@@ -82,10 +82,11 @@ libpq options, and approval scope in one place.
 External comparison extensions can use the same setup surface:
 
 ```sh
-/Users/peter/.cargo/bin/ecaz dev install pgvector --pg 18
-/Users/peter/.cargo/bin/ecaz dev install vectorscale \
+${HOME}/.cargo/bin/ecaz dev install pgvector --pg 18 \
+  --repo "${PGVECTOR_REPO}"
+${HOME}/.cargo/bin/ecaz dev install vectorscale \
   --pg 18 \
-  --repo "$HOME/dev_bak/pgvectorscale/pgvectorscale" \
+  --repo "${PGVECTORSCALE_REPO}" \
   --cargo-pgrx /tmp/pgvectorscale-cargo-pgrx-0.16.1/bin/cargo-pgrx
 ```
 
diff --git a/crates/ecaz-cli/src/commands/dev/install.rs b/crates/ecaz-cli/src/commands/dev/install.rs
index 5a7a269782..cc4adbe3c8 100644
--- a/crates/ecaz-cli/src/commands/dev/install.rs
+++ b/crates/ecaz-cli/src/commands/dev/install.rs
@@ -50,7 +50,7 @@ pub struct InstallEcazPgTestArgs {
 #[derive(Args, Debug)]
 pub struct InstallPgvectorArgs {
     /// pgvector repository checkout.
-    #[arg(long, env = "PGVECTOR_REPO", default_value_os_t = default_pgvector_repo())]
+    #[arg(long, env = "PGVECTOR_REPO")]
     repo: PathBuf,
 
     /// PostgreSQL major version to install against.
@@ -69,11 +69,7 @@ pub struct InstallPgvectorArgs {
 #[derive(Args, Debug)]
 pub struct InstallVectorscaleArgs {
     /// pgvectorscale extension crate checkout.
-    #[arg(
-        long,
-        env = "PGVECTORSCALE_REPO",
-        default_value_os_t = default_vectorscale_repo()
-    )]
+    #[arg(long, env = "PGVECTORSCALE_REPO")]
     repo: PathBuf,
 
     /// cargo-pgrx binary to use for the install.
@@ -284,14 +280,6 @@ fn home_dir() -> PathBuf {
         .unwrap_or_else(|| PathBuf::from("/"))
 }
 
-fn default_pgvector_repo() -> PathBuf {
-    home_dir().join("dev_bak/pgvector")
-}
-
-fn default_vectorscale_repo() -> PathBuf {
-    home_dir().join("dev_bak/pgvectorscale/pgvectorscale")
-}
-
 fn default_cargo_pgrx_bin() -> PathBuf {
     home_dir().join(".cargo/bin/cargo-pgrx")
 }
diff --git a/plan/design/spire-phase11-paper-parity-production-gate.md b/plan/design/spire-phase11-paper-parity-production-gate.md
index a6b51e31e8..9293aa8240 100644
--- a/plan/design/spire-phase11-paper-parity-production-gate.md
+++ b/plan/design/spire-phase11-paper-parity-production-gate.md
@@ -2,7 +2,7 @@
 
 Status: active Phase 11.1 gate
 Task: Task 30 Phase 11
-Paper basis: `/home/peter/dev_bak/papers/2512.17264v1.pdf`
+Paper basis: arXiv:2512.17264v1
 
 ## Source Basis
 
diff --git a/plan/tasks/240-artifact-path-sanitization.md b/plan/tasks/240-artifact-path-sanitization.md
new file mode 100644
index 0000000000..61d929f674
--- /dev/null
+++ b/plan/tasks/240-artifact-path-sanitization.md
@@ -0,0 +1,42 @@
+# Task 240: Artifact Path Sanitization
+
+Status: **implementation in progress** (2026-08-29). Priority: P0 privacy and
+evidence hygiene. Origin: private Engineering Assurance gap audit.
+
+## Why
+
+The public default branch contains machine-specific workstation paths in authored
+configuration and historical evidence. GitHub code search exposed only a subset; a
+local tracked-file audit found 6,396 affected files. The values disclose local layout,
+are non-portable, and can accidentally identify rights-unreviewed source locations.
+
+## Goal
+
+Remove the explicitly identified internal scratch locator from the current tree, make
+external-source references durable, and prevent any newly added workstation path from
+reaching GitHub without rewriting historical benchmark evidence blindly.
+
+## Scope
+
+1. Replace current authored internal-source locators with durable public identifiers or
+   explicit operator inputs.
+2. Remove machine-specific default comparison-repository locations from `ecaz dev
+   install`; require `--repo` or the documented environment variable.
+3. Add a diff-aware guard that detects Unix, root, and Windows workstation home paths,
+   reports only file/line/category, and never repeats the matched value in CI logs.
+4. Run the guard on every pull request and main-branch push in a lightweight workflow.
+5. Preserve a transparent distinction between current-tree redaction and historical Git
+   reachability. Do not rewrite history or silently change digest-bound evidence.
+
+## Acceptance
+
+1. The current tree contains no reference to the internal scratch locator identified by
+   the audit.
+2. Synthetic tests prove added Unix, root, and Windows workstation paths fail while
+   removals and portable environment-based paths pass.
+3. Failure output contains the file, line, and category but not the sensitive value.
+4. The CLI comparison installers have no machine-specific repository default.
+5. The lightweight GitHub workflow enforces the guard without invoking benchmark or
+   PostgreSQL lanes.
+6. A review packet records the legacy-corpus count, redaction boundary, test evidence,
+   and the separate unresolved history-remediation decision.
diff --git a/plan/tasks/87-candidate-batched-scoring-across-ams.md b/plan/tasks/87-candidate-batched-scoring-across-ams.md
index 787437e502..b3e5a5a04a 100644
--- a/plan/tasks/87-candidate-batched-scoring-across-ams.md
+++ b/plan/tasks/87-candidate-batched-scoring-across-ams.md
@@ -379,7 +379,7 @@ fix, not Task 87.
   the calibration prototype already produces `Prepared…Query`
   shapes that batch naturally.
 - **pgvectorscale** is a read-only reference: clone at
-  `/Users/peter/dev_bak/pgvectorscale/`; key files
+  the `timescale/pgvectorscale` repository; key files
   `access_method/scan.rs` (resort_buffer pattern) and
   `access_method/graph/mod.rs` (streaming iteration).
 
@@ -426,7 +426,7 @@ fix, not Task 87.
 - Task 86 packet 001 transferability matrix (per-AM block-
   kernel fit ranking)
 - pgvectorscale resort_buffer pattern:
-  `/Users/peter/dev_bak/pgvectorscale/pgvectorscale/src/access_method/scan.rs`
+  `pgvectorscale/src/access_method/scan.rs`
 - FR-038 (benchmark provenance): every suite checked-in JSON
 - ADR-075 (Task 65b stepping stone framing) — similar
   staged-rollout pattern
diff --git a/plan/tasks/88-streaming-ann-result-iteration.md b/plan/tasks/88-streaming-ann-result-iteration.md
index 49b828b13b..e660707794 100644
--- a/plan/tasks/88-streaming-ann-result-iteration.md
+++ b/plan/tasks/88-streaming-ann-result-iteration.md
@@ -8,7 +8,7 @@ Priority: 2 (hybrid-search support; follows Task 87)
 
 Task 86 investigation surfaced pgvectorscale's
 `StreamingDiskANN` pattern (see
-`/Users/peter/dev_bak/pgvectorscale/pgvectorscale/src/access_method/graph/mod.rs`
+`timescale/pgvectorscale:pgvectorscale/src/access_method/graph/mod.rs`
 `greedy_search_streaming_init` + `greedy_search_iterate`).
 The pattern yields candidates lazily to the PostgreSQL
 executor in approximate-score order, with a resort buffer +
@@ -200,7 +200,7 @@ the two.
   the post-filter use case justification — don't expand
   Task 88's scope.
 - **pgvectorscale** is the reference implementation
-  (`/Users/peter/dev_bak/pgvectorscale/`). Key files:
+  (`timescale/pgvectorscale`). Key files:
   - `pgvectorscale/src/access_method/scan.rs`
     (`resort_buffer`, `StreamingStats`)
   - `pgvectorscale/src/access_method/graph/mod.rs`
@@ -243,7 +243,7 @@ the two.
   (surfaced the streaming-vs-batching distinction)
 - Task 87 (predecessor): `plan/tasks/87-candidate-batched-scoring-across-ams.md`
 - pgvectorscale `StreamingDiskANN` reference:
-  `/Users/peter/dev_bak/pgvectorscale/`
+  `timescale/pgvectorscale`
 - pgvectorscale README streaming section: *"The
   post-filtering implementation, while slower, is streaming
   and correct…"*
diff --git a/plan/tasks/89-turboquant-tqplus-cross-am-validation.md b/plan/tasks/89-turboquant-tqplus-cross-am-validation.md
index 183216f2f6..ed73f43a86 100644
--- a/plan/tasks/89-turboquant-tqplus-cross-am-validation.md
+++ b/plan/tasks/89-turboquant-tqplus-cross-am-validation.md
@@ -293,7 +293,7 @@ permission to skip the AM.
   (includes reviewer feedback on the original closeout)
 - Reverted Task 86 commits (TQ+ work, preserved in git
   history): `e0ae9fe7d`, `c7e85e8ac`, `16f1e6104`, etc.
-- pgvectorscale reference: `/Users/peter/dev_bak/pgvectorscale/`
+- pgvectorscale reference: `timescale/pgvectorscale`
 
 ## Estimated size
 
diff --git a/plan/tasks/README.md b/plan/tasks/README.md
index a9443567d9..8e7ae62171 100644
--- a/plan/tasks/README.md
+++ b/plan/tasks/README.md
@@ -279,6 +279,8 @@ These pre-lane task files are retained only for historical context under
 238. `238-ec-distann-retry-snapshot-use-after-free.md` — **implementation and evidence merged; packet 001 outside closeout review-open** (updated 2026-08-27, P0 correctness): the equivalent `RegisteredSnapshotGuard` lifetime fix remains on main via `15f7fcf5f` / Task 167 PR #77. Checkpoint `7d4103885` restores the deterministic forced-retry regression and explicit loopback-only plaintext test conninfo; both PG18 callers pass on base `3c81319a3`. PR #91 merged the implementation and evidence as `bcc9b0bfa`; only the outside ACCEPT remains open. Evidence and request: `reviews/task-238/001-retry-snapshot-uaf/`.
 239. `239-ec-distann-bounded-read-overfetch.md` — **complete — review-closed ACCEPT; HARNESS REGRESSION CORRECTED, EXACT-MAIN LAZY-10 SEMANTIC PATH RESTORED TO 10/10; no rerun; packet 004 10k/50k/100k NOT triggered; Task 229 / 230--233 semantic-surface blocker LIFTED** (updated 2026-08-26, P1 correctness/performance): packet 001's 12/10 result was a shared-session batch-size GUC leak in the benchmark harness — a variant at batch size 10 emitted no `SET`, so the "lazy-10" arm inherited the eager control's 0 — not production overfetch, and the bound was never widened. Packet 002 corrected the harness but its cross-SHA runner aborted pre-semantics on 40-versus-37 stage rows. Packet 003 removed that confounder with exact-main runner/extension plus only the harness fix; its sole C1--C5 run passed every preregistered gate — nine scenarios exactly once in both logs, seven core rows at `control_batch_size=0 candidate_batch_size=10` with exact identity and zero duplicates, `exactly_one_window` 6 remote + 4 local = 10 reads/bound 10, mixed and outage pass, routed DELETE+VACUUM pass, both recall arms 0.9990 over 200 queries / 2,000 trials with predictions byte-identical to packet 001 — and the stopped 1.2 GB fixture was removed after capture. `git diff 41392c011 def565270 -- src` is empty, so no production runtime behavior changed and the 10k/50k/100k closeout matrix does not apply. Determinism rests on the same 6/4/10 split and digests at three independent extension SHAs (task-191 `7883cfcf`, task-198 `2ff72b3e`, this run `4ab2aa9a9`), so no rerun was needed. Carried follow-ups for whichever task next touches the multinode semantic harness: mixed/outage failures emit no structured `pass=false` row; `owner_payload_plan_cache` retains the same conditional-`SET` leak shape; recall/latency child batch size stays inferred. Closeout verdict: `reviews/task-239/003-main-baseline-semantic-proof/feedback/2026-08-26-02-reviewer.md`; result decision: `reviews/task-239/003-main-baseline-semantic-proof/artifacts/live-run-decision.md`; review request: `reviews/task-239/003-main-baseline-semantic-proof/request.md`; live authorization: `reviews/task-239/003-main-baseline-semantic-proof/feedback/2026-08-26-01-reviewer.md`.
 
+240. `240-artifact-path-sanitization.md` — **implementation in progress** (2026-08-29, P0 privacy/evidence hygiene): remove current-tree internal scratch locators, require portable comparison-repository inputs, and add a non-echoing diff guard for newly introduced workstation paths; historical Git remediation remains a separate explicit decision.
+
 ## Coordination rules
 
 - Freeze binary datum layout before downstream work expands.
diff --git a/reviews/task-08/014-205-a4-quantizer-impl-mismatch-audit/request.md b/reviews/task-08/014-205-a4-quantizer-impl-mismatch-audit/request.md
index f0b4815e94..db2912237d 100644
--- a/reviews/task-08/014-205-a4-quantizer-impl-mismatch-audit/request.md
+++ b/reviews/task-08/014-205-a4-quantizer-impl-mismatch-audit/request.md
@@ -1,7 +1,7 @@
 # Review Request: A4 Quantizer Implementation Mismatch Audit
 
 Basis: `main` working tree after review `204`, plus upstream reference at
-`~/dev_bak/TurboQuantDB/`
+`https://github.com/jyunming/TurboQuantDB`
 
 ## Why This Packet Exists
 
diff --git a/reviews/task-16/001-204-a4-full-vs-tiled-fwht-turboquantdb/feedback/2026-04-09-01-reviewer.md b/reviews/task-16/001-204-a4-full-vs-tiled-fwht-turboquantdb/feedback/2026-04-09-01-reviewer.md
index 8a76bc7c7e..154a1fbc30 100644
--- a/reviews/task-16/001-204-a4-full-vs-tiled-fwht-turboquantdb/feedback/2026-04-09-01-reviewer.md
+++ b/reviews/task-16/001-204-a4-full-vs-tiled-fwht-turboquantdb/feedback/2026-04-09-01-reviewer.md
@@ -16,14 +16,15 @@ Request:
 
 ## Additional Reference: prism-mcp TurboQuant Implementation
 
-A third independent TurboQuant implementation exists in `~/dev_bak/prism-mcp/`. It is a pure
+A third independent TurboQuant implementation exists in
+`https://github.com/dcostenco/prism-mcp`. It is a pure
 TypeScript port used for Prism MCP's "Quantized Agentic Memory" feature (v5.0). Comparing it
 alongside TurboQuantDB strengthens several conclusions from review 204 and surfaces one new
 finding.
 
 ### prism-mcp Architecture Summary
 
-Source: `~/dev_bak/prism-mcp/src/utils/turboquant.ts` (~840 lines)
+Source: `dcostenco/prism-mcp:src/utils/turboquant.ts` (~840 lines)
 
 - **Rotation**: QR decomposition (Householder) of a random Gaussian matrix — NOT FWHT/SRHT
 - **Dimension**: 768 (Gemini text-embedding-004), no power-of-2 constraint
@@ -150,7 +151,7 @@ quality lever.
 
 ## Paper Reference: Rotation Is QR, Not FWHT
 
-Source: `~/dev_bak/turboquant-2504.19874.pdf` (arXiv:2504.19874, ICLR 2026)
+Source: arXiv:2504.19874 (ICLR 2026)
 
 The TurboQuant paper specifies QR decomposition, not FWHT:
 
@@ -216,7 +217,7 @@ contract within tqvector's SRHT-based architecture.
 
 ## TurboQuantDB Published Recall Benchmarks
 
-Source: `~/dev_bak/TurboQuantDB/README.md`, section "Benchmarks"
+Source: `jyunming/TurboQuantDB:README.md`, section "Benchmarks"
 
 TurboQuantDB publishes recall numbers on **real embeddings** (50k × 1536, DBpedia OpenAI
 embeddings, top_k=10):
@@ -245,7 +246,8 @@ is already good enough that the residual is small and the 1-bit correction is ef
 
 ## TurboQuantDB Deep-Dive: Code-Level Architecture
 
-Full source read of `~/dev_bak/TurboQuantDB/src/quantizer/` and `src/linalg/hadamard.rs`.
+Full source read of `jyunming/TurboQuantDB:src/quantizer/` and
+`src/linalg/hadamard.rs`.
 
 ### README Describes QR, Code Uses SRHT
 
diff --git a/reviews/task-16/001-204-a4-full-vs-tiled-fwht-turboquantdb/request.md b/reviews/task-16/001-204-a4-full-vs-tiled-fwht-turboquantdb/request.md
index 0d3cf20422..3864abb828 100644
--- a/reviews/task-16/001-204-a4-full-vs-tiled-fwht-turboquantdb/request.md
+++ b/reviews/task-16/001-204-a4-full-vs-tiled-fwht-turboquantdb/request.md
@@ -1,6 +1,7 @@
 # Review Request: A4 Full vs Tiled FWHT — TurboQuantDB Comparison
 
-Basis: `main` working tree, with reference to `~/dev_bak/TurboQuantDB/`
+Basis: `main` working tree, with reference to
+`https://github.com/jyunming/TurboQuantDB`
 
 ## Summary
 
diff --git a/reviews/task-16/054-363-c1-adr030-v2-pgvector-size-and-runtime-baseline/request.md b/reviews/task-16/054-363-c1-adr030-v2-pgvector-size-and-runtime-baseline/request.md
index becbc2ddc0..d6d2a242c8 100644
--- a/reviews/task-16/054-363-c1-adr030-v2-pgvector-size-and-runtime-baseline/request.md
+++ b/reviews/task-16/054-363-c1-adr030-v2-pgvector-size-and-runtime-baseline/request.md
@@ -53,7 +53,7 @@ Updated:
 Concrete changes:
 
 1. added a reproducible scratch install wrapper for a local pgvector checkout
-   (`/home/peter/dev_bak/pgvector` by default)
+   (the checkout supplied through `PGVECTOR_REPO` or `--repo`)
 2. the script installs pgvector into the same pg17 pgrx/Postgres tree already
    used by the scratch cluster:
    - `PG_CONFIG=/home/peter/.pgrx/17.9/pgrx-install/bin/pg_config`
@@ -195,4 +195,3 @@ The next useful follow-up is not more size work. It is one of:
    isolated grouped index without the special verified harness
 3. make the product call explicitly: latency-first compressed ANN versus
    higher-recall full-vector ANN
-
diff --git a/reviews/task-203/001-decision-reaudit/artifacts/manifest.md b/reviews/task-203/001-decision-reaudit/artifacts/manifest.md
index a335b0b0c5..6276d5e0b1 100644
--- a/reviews/task-203/001-decision-reaudit/artifacts/manifest.md
+++ b/reviews/task-203/001-decision-reaudit/artifacts/manifest.md
@@ -24,7 +24,7 @@ existing packets, task files, spec files, or source.
 
 `DISTRIBUTEDANN: Efficient Scaling of a Single DISKANN Graph Across Thousands of
 Computers`, arXiv:2509.06046v1, 7 Sep 2025. Local copy:
-`~/dev_bak/papers/distributedann-2509.06046.pdf` (8 pages). Sections cited:
+DistributedANN, arXiv:2509.06046 (8 pages). Sections cited:
 §2.2 (index layout modifications, head index), §2.3 (near-data computation,
 Algorithm 1), §2.4 (orchestration service, Algorithm 2), §3 (graph
 construction), §4 (evaluation parameters), §4.1 (scaling), §4.2 (reliability).
diff --git a/reviews/task-204/001-arm-fidelity/feedback/2026-07-30-02-reviewer.md b/reviews/task-204/001-arm-fidelity/feedback/2026-07-30-02-reviewer.md
index 7df0172a39..c3e5d0cce1 100644
--- a/reviews/task-204/001-arm-fidelity/feedback/2026-07-30-02-reviewer.md
+++ b/reviews/task-204/001-arm-fidelity/feedback/2026-07-30-02-reviewer.md
@@ -12,7 +12,7 @@ Short addendum to `2026-07-30-01-reviewer.md`, narrowing one instruction in it.
 
 My seq-01 "Open item 1" asked for two things under Phase 3: a computed
 cross-scale growth row, and a gate that fails the run on breach. After reading
-`~/dev_bak/papers/distributedann-2509.06046.pdf` directly, I want to split those
+DistributedANN (arXiv:2509.06046) directly, I want to split those
 apart. Full reasoning and citations are in
 `reviews/task-205/003-ab/feedback/2026-07-30-02-reviewer.md` and
 `reviews/task-205/003-ab/artifacts/paper-algorithm-citations.md`.
diff --git a/reviews/task-205/001-contract/feedback/2026-07-30-02-reviewer.md b/reviews/task-205/001-contract/feedback/2026-07-30-02-reviewer.md
index 4847deba11..ea5df63074 100644
--- a/reviews/task-205/001-contract/feedback/2026-07-30-02-reviewer.md
+++ b/reviews/task-205/001-contract/feedback/2026-07-30-02-reviewer.md
@@ -9,7 +9,7 @@ seq: 02
 # Task 205 / 001-contract — the paper's actual contract values
 
 Addendum to `2026-07-30-01-reviewer.md`. I read the paper directly
-(`~/dev_bak/papers/distributedann-2509.06046.pdf`) and transcribed the two
+(DistributedANN, arXiv:2509.06046) and transcribed the two
 algorithms verbatim into
 `reviews/task-205/003-ab/artifacts/paper-algorithm-citations.md` so these
 citations are checkable from inside the packet.
diff --git a/reviews/task-205/003-ab/artifacts/paper-algorithm-citations.md b/reviews/task-205/003-ab/artifacts/paper-algorithm-citations.md
index bf011ea087..79ec3a9dc3 100644
--- a/reviews/task-205/003-ab/artifacts/paper-algorithm-citations.md
+++ b/reviews/task-205/003-ab/artifacts/paper-algorithm-citations.md
@@ -2,7 +2,7 @@
 
 Source: `DISTRIBUTEDANN: Efficient Scaling of a Single DISKANN Graph Across
 Thousands of Computers`, arXiv:2509.06046, 8 pages. Local copy read at
-`~/dev_bak/papers/distributedann-2509.06046.pdf`. Text extracted with `pypdf`;
+DistributedANN (arXiv:2509.06046). Text extracted with `pypdf`;
 line breaks and hyphenation are as extracted.
 
 These are transcribed here so the review's citations are checkable from inside
diff --git a/reviews/task-205/003-ab/feedback/2026-07-30-02-reviewer.md b/reviews/task-205/003-ab/feedback/2026-07-30-02-reviewer.md
index 992289ffcb..97ddbbab61 100644
--- a/reviews/task-205/003-ab/feedback/2026-07-30-02-reviewer.md
+++ b/reviews/task-205/003-ab/feedback/2026-07-30-02-reviewer.md
@@ -9,7 +9,7 @@ seq: 02
 # Task 205 / 003-ab — paper grounding for the re-run and the NFR-021 verdict
 
 Addendum to `2026-07-30-01-reviewer.md`, after reading
-`~/dev_bak/papers/distributedann-2509.06046.pdf` directly. Verbatim algorithm and
+DistributedANN (arXiv:2509.06046) directly. Verbatim algorithm and
 results text is transcribed in `artifacts/paper-algorithm-citations.md`.
 
 Both seq-01 P1s survive the reading, and both get sharper. One correction to my
diff --git a/reviews/task-25/014-20013-task25-symphony-prerequisite-finding/request.md b/reviews/task-25/014-20013-task25-symphony-prerequisite-finding/request.md
index bf02225aee..5c8dc22701 100644
--- a/reviews/task-25/014-20013-task25-symphony-prerequisite-finding/request.md
+++ b/reviews/task-25/014-20013-task25-symphony-prerequisite-finding/request.md
@@ -12,9 +12,9 @@ the task-25 closing posture and the task-27 start date:
    wrong configuration for Symphony.
 
 Artifacts:
-- `~/dev_bak/papers/symphonyqg-2025-sigmod-arxiv-2411.12229.pdf`
-- `~/dev_bak/papers/rabitq-2024-sigmod-arxiv-2405.12497.pdf`
-- `~/dev_bak/papers/extended-rabitq-2025-sigmod-arxiv-2409.09913.pdf`
+- SymphonyQG, arXiv:2411.12229
+- RaBitQ, arXiv:2405.12497
+- Extended RaBitQ, arXiv:2409.09913
 
 Branch: `task25-rabitq-stage1-phase0` (slice 14 builds on `3431476`).
 
@@ -137,7 +137,7 @@ trait surface exactly as today.
   subsection to the same section with the API shape, the
   equation-6 decomposition, and the rationale for keeping the
   centered path off the `Quantizer` trait.
-- Papers archived in `~/dev_bak/papers/` (not committed to repo
+- Papers inspected from local authorized copies (not committed to repo
   since outside the project tree and large binary).
 
 ## What this slice does NOT do
diff --git a/reviews/task-25/015-20014-task25-centered-api/feedback/2026-04-23-01-reviewer.md b/reviews/task-25/015-20014-task25-centered-api/feedback/2026-04-23-01-reviewer.md
index bad56eca16..a61b3cf883 100644
--- a/reviews/task-25/015-20014-task25-centered-api/feedback/2026-04-23-01-reviewer.md
+++ b/reviews/task-25/015-20014-task25-centered-api/feedback/2026-04-23-01-reviewer.md
@@ -65,7 +65,7 @@ seq: 01
 ## Validation
 
 - Read the task-25 and task-27 docs, the slice-6/14/15 packets, and ADR-045.
-- Verified the Symphony paper passages locally from `~/dev_bak/papers/`.
+- Verified the Symphony paper passages from an authorized local copy.
 - `cargo test`
 - `cargo check -p ecaz-cli`
 
diff --git a/reviews/task-29/014-11095-task29-diskann-pgvectorscale-comparison/feedback.md b/reviews/task-29/014-11095-task29-diskann-pgvectorscale-comparison/feedback.md
index 17660bcb8c..42e0c9efaf 100644
--- a/reviews/task-29/014-11095-task29-diskann-pgvectorscale-comparison/feedback.md
+++ b/reviews/task-29/014-11095-task29-diskann-pgvectorscale-comparison/feedback.md
@@ -7,7 +7,7 @@ Recipient: coder-1
 This is a code-level evaluation of `src/am/ec_diskann/` and a comparison to
 pgvectorscale's Vamana implementation. It is informed by reading every probe
 packet `11087..11094` and the relevant ec_diskann source plus
-`~/dev_bak/pgvectorscale/pgvectorscale/src/access_method/` (cloned at HEAD).
+`timescale/pgvectorscale:pgvectorscale/src/access_method/` (cloned at HEAD).
 
 The review excludes pgvectorscale's streaming-iterator amgettuple shape per
 the request; everything else (graph build, prune, search, distance handling,
diff --git a/reviews/task-30/581-30658-spire-phase9-routing-plan/feedback/2026-05-09-01-reviewer.md b/reviews/task-30/581-30658-spire-phase9-routing-plan/feedback/2026-05-09-01-reviewer.md
index 0be6ca41fe..9cebef3805 100644
--- a/reviews/task-30/581-30658-spire-phase9-routing-plan/feedback/2026-05-09-01-reviewer.md
+++ b/reviews/task-30/581-30658-spire-phase9-routing-plan/feedback/2026-05-09-01-reviewer.md
@@ -11,7 +11,7 @@ seq: 01
 Review pass on `661d5008`, focused on graph structure, composition across
 hierarchy levels, multi-NVMe / multi-instance behavior, the query path, and
 the current custom libpq surfaces. Paper cross-check was against
-`/home/peter/dev_bak/papers/2512.17264v1.pdf`, which describes SPIRE as a
+arXiv:2512.17264v1, which describes SPIRE as a
 hierarchical index with top/root graph routing, level-by-level top-m descent,
 boundary replication, disaggregated index storage, and stateless query engines
 that fetch relevant index data from the store.
diff --git a/reviews/task-30/613-30691-spire-phase11-production-parity-plan/artifacts/paper-parity-seed.md b/reviews/task-30/613-30691-spire-phase11-production-parity-plan/artifacts/paper-parity-seed.md
index e47e5ec566..3ddcc6fe15 100644
--- a/reviews/task-30/613-30691-spire-phase11-production-parity-plan/artifacts/paper-parity-seed.md
+++ b/reviews/task-30/613-30691-spire-phase11-production-parity-plan/artifacts/paper-parity-seed.md
@@ -4,7 +4,7 @@ Reference basis:
 
 - Paper cross-check from
   `reviews/task-30/581-30658-spire-phase9-routing-plan/feedback/2026-05-09-01-reviewer.md`
-  against `/home/peter/dev_bak/papers/2512.17264v1.pdf`.
+  against arXiv:2512.17264v1.
 - Phase 9 and 10 local architecture closeout in:
   - `plan/tasks/task30-phase9-spire-graph-architecture.md`
   - `plan/tasks/task30-phase10-spire-execution-performance.md`
diff --git a/reviews/task-30/614-30692-spire-phase11-paper-parity-gate/artifacts/paper-source-outline.md b/reviews/task-30/614-30692-spire-phase11-paper-parity-gate/artifacts/paper-source-outline.md
index 052a6bd6dd..5872d6453e 100644
--- a/reviews/task-30/614-30692-spire-phase11-paper-parity-gate/artifacts/paper-source-outline.md
+++ b/reviews/task-30/614-30692-spire-phase11-paper-parity-gate/artifacts/paper-source-outline.md
@@ -1,6 +1,6 @@
 # SPire Paper Source Outline
 
-Source file: `/home/peter/dev_bak/papers/2512.17264v1.pdf`
+Source: arXiv:2512.17264v1
 
 The local PDF metadata identifies the paper as "Scalable Distributed Vector
 Search via Accuracy Preserving Index Construction". The decoded outline used
diff --git a/reviews/task-86/001-turbovec-tq-analysis/artifacts/manifest.md b/reviews/task-86/001-turbovec-tq-analysis/artifacts/manifest.md
index 2266a11456..87a46ba412 100644
--- a/reviews/task-86/001-turbovec-tq-analysis/artifacts/manifest.md
+++ b/reviews/task-86/001-turbovec-tq-analysis/artifacts/manifest.md
@@ -12,7 +12,7 @@
   - `sed` / `rg` / `nl` source inspection against local TurboVec and current
     repo sources
 - TurboVec source snapshot:
-  - Local path: `/Users/peter/dev_bak/turbovec`
+  - Source identity: TurboVec repository snapshot
   - SHA: `efe29a184986cbf562a9847c2ac52a2990bfaca2`
   - Status: clean `main`
 
diff --git a/reviews/task-86/001-turbovec-tq-analysis/artifacts/turbovec-tq-analysis.md b/reviews/task-86/001-turbovec-tq-analysis/artifacts/turbovec-tq-analysis.md
index be01a6d7aa..0f7c2d76c6 100644
--- a/reviews/task-86/001-turbovec-tq-analysis/artifacts/turbovec-tq-analysis.md
+++ b/reviews/task-86/001-turbovec-tq-analysis/artifacts/turbovec-tq-analysis.md
@@ -4,7 +4,7 @@ Scope: compare TurboVec's TurboQuant implementation to our TurboQuant
 implementation only. This report intentionally does not evaluate other
 quantizers or external leaderboards.
 
-TurboVec source inspected: `/Users/peter/dev_bak/turbovec` at
+TurboVec source inspected at
 `efe29a184986cbf562a9847c2ac52a2990bfaca2`.
 
 Our source inspected: `/Users/peter/dev/tqvector` at
@@ -30,17 +30,17 @@ build lookup tables. The query is not packed into the same database code format.
 
 Relevant source:
 
-- `/Users/peter/dev_bak/turbovec/turbovec/src/lib.rs:96` defines
+- `TurboVec@efe29a1:turbovec/src/lib.rs:96` defines
   `TurboQuantIndex` with `packed_codes`, `scales`, TQ+ shift/scale arrays, and
   lazy blocked-code caches.
-- `/Users/peter/dev_bak/turbovec/turbovec/src/lib.rs:250` through `:283`
+- `TurboVec@efe29a1:turbovec/src/lib.rs:250` through `:283`
   initializes rotation/codebook state and calls `encode::encode`.
-- `/Users/peter/dev_bak/turbovec/turbovec/src/encode.rs:1` through `:27`
+- `TurboVec@efe29a1:turbovec/src/encode.rs:1` through `:27`
   documents normalize, rotate, TQ+ calibrate, quantize, pack, and scale.
-- `/Users/peter/dev_bak/turbovec/turbovec/src/encode.rs:65` through `:133`
+- `TurboVec@efe29a1:turbovec/src/encode.rs:65` through `:133`
   implements normalize, dense rotation, calibration, packed-code allocation,
   and per-row fused quantize/scale/pack.
-- `/Users/peter/dev_bak/turbovec/turbovec/src/encode.rs:245` through `:352`
+- `TurboVec@efe29a1:turbovec/src/encode.rs:245` through `:352`
   explains and implements the per-vector renormalization scalar.
 
 Encoding shape:
@@ -61,16 +61,16 @@ the same calibration, so all vectors live in one calibrated coordinate system.
 
 Relevant source:
 
-- `/Users/peter/dev_bak/turbovec/turbovec/src/search.rs:1466` through `:1530`
+- `TurboVec@efe29a1:turbovec/src/search.rs:1466` through `:1530`
   implements full query search setup.
-- `/Users/peter/dev_bak/turbovec/turbovec/src/search.rs:1492` through `:1508`
+- `TurboVec@efe29a1:turbovec/src/search.rs:1492` through `:1508`
   rotates all queries with a batched GEMM.
-- `/Users/peter/dev_bak/turbovec/turbovec/src/search.rs:1510` through `:1518`
+- `TurboVec@efe29a1:turbovec/src/search.rs:1510` through `:1518`
   applies inverse TQ+ calibration.
-- `/Users/peter/dev_bak/turbovec/turbovec/src/search.rs:1419` through `:1455`
+- `TurboVec@efe29a1:turbovec/src/search.rs:1419` through `:1455`
   defines `q_calib[d] = q_rot[d] / scale_tq[d]` and bias
   `-sum(q_rot[d] * shift[d])`.
-- `/Users/peter/dev_bak/turbovec/turbovec/src/search.rs:1165` through `:1280`
+- `TurboVec@efe29a1:turbovec/src/search.rs:1165` through `:1280`
   builds per-query `u8` nibble lookup tables with per-subtable mins, one shared
   scale, and a bias.
 
@@ -83,20 +83,20 @@ decompressed into dense float vectors at query time.
 
 Relevant source:
 
-- `/Users/peter/dev_bak/turbovec/turbovec/src/lib.rs:357` through `:475`
+- `TurboVec@efe29a1:turbovec/src/lib.rs:357` through `:475`
   exposes `search` and `search_with_mask`, materializes a blocked-code cache,
   packs optional filters into a slot bitset, and calls `search::search`.
-- `/Users/peter/dev_bak/turbovec/turbovec/src/pack.rs:1` through `:60`
+- `TurboVec@efe29a1:turbovec/src/pack.rs:1` through `:60`
   repacks per-vector bit-plane codes into SIMD-blocked layout.
-- `/Users/peter/dev_bak/turbovec/turbovec/src/pack.rs:62` through `:88`
+- `TurboVec@efe29a1:turbovec/src/pack.rs:62` through `:88`
   uses an x86 FAISS-style 32-vector block layout with split hi/lo nibbles.
-- `/Users/peter/dev_bak/turbovec/turbovec/src/pack.rs:90` through `:114`
+- `TurboVec@efe29a1:turbovec/src/pack.rs:90` through `:114`
   uses a sequential 32-vector block layout on non-x86.
-- `/Users/peter/dev_bak/turbovec/turbovec/src/search.rs:1293` through `:1335`
+- `TurboVec@efe29a1:turbovec/src/search.rs:1293` through `:1335`
   implements block and block-pair filter skips.
-- `/Users/peter/dev_bak/turbovec/turbovec/src/search.rs:1337` through `:1417`
+- `TurboVec@efe29a1:turbovec/src/search.rs:1337` through `:1417`
   shows scalar scoring over all blocks and lanes with fused heap maintenance.
-- `/Users/peter/dev_bak/turbovec/turbovec/src/search.rs:1532` onward dispatches
+- `TurboVec@efe29a1:turbovec/src/search.rs:1532` onward dispatches
   platform-specific scoring and top-k.
 
 There is no routing layer in the inspected implementation:
@@ -118,10 +118,10 @@ candidate generation, graph traversal, page layout, heap rerank, or tuple I/O.
 
 TurboVec file layout source:
 
-- `/Users/peter/dev_bak/turbovec/turbovec/src/io.rs:178` through `:213` writes
+- `TurboVec@efe29a1:turbovec/src/io.rs:178` through `:213` writes
   bit width, dimension, vector count, packed codes, per-vector scales, and TQ+
   calibration.
-- `/Users/peter/dev_bak/turbovec/turbovec/src/io.rs:262` through `:276` reads
+- `TurboVec@efe29a1:turbovec/src/io.rs:262` through `:276` reads
   `packed_bytes = (dim / 8) * bit_width * n_vectors` and `n_vectors` f32
   scales.
 
diff --git a/reviews/task-87/001-phase1-design/artifacts/source-scoring-map.md b/reviews/task-87/001-phase1-design/artifacts/source-scoring-map.md
index 82ee26f8a7..37374a117e 100644
--- a/reviews/task-87/001-phase1-design/artifacts/source-scoring-map.md
+++ b/reviews/task-87/001-phase1-design/artifacts/source-scoring-map.md
@@ -126,7 +126,7 @@ Task 87 mapping:
 
 Reference file:
 
-- `/home/peter/dev_bak/pgvectorscale/pgvectorscale/src/access_method/scan.rs`
+- `timescale/pgvectorscale:pgvectorscale/src/access_method/scan.rs`
 
 Relevant resort-buffer lines:
 
diff --git a/scripts/check_workstation_paths.py b/scripts/check_workstation_paths.py
new file mode 100644
index 0000000000..6c64bc5c65
--- /dev/null
+++ b/scripts/check_workstation_paths.py
@@ -0,0 +1,136 @@
+#!/usr/bin/env python3
+"""Reject newly added machine-specific workstation paths without echoing them."""
+
+from __future__ import annotations
+
+import argparse
+import re
+import subprocess
+import sys
+from dataclasses import dataclass
+
+
+HUNK = re.compile(r"^@@ -\d+(?:,\d+)? \+(\d+)(?:,\d+)? @@")
+PATH_PATTERNS = (
+    ("tilde workstation home path", re.compile(r"(?:^|(?<=\s)|(?<=[\"'`]))~/(?=\S)")),
+    (
+        "Unix workstation home path",
+        re.compile(r"/(?:home|Users)/[^/\s\"'`]+(?=/|\s|[\"'`]|$)"),
+    ),
+    ("root workstation path", re.compile(r"/root(?=/|\s|[\"'`]|$)")),
+    (
+        "Windows workstation home path",
+        re.compile(
+            r"(?i)\b[A-Z]:[\\/](?:Users|Documents and Settings)[\\/]"
+            r"[^\\/\s\"'`]+(?=[\\/]|\s|[\"'`]|$)"
+        ),
+    ),
+)
+
+
+@dataclass(frozen=True)
+class AddedLine:
+    path: str
+    line_number: int
+    text: str
+
+
+@dataclass(frozen=True)
+class Violation:
+    path: str
+    line_number: int
+    label: str
+
+
+def parse_added_lines(diff: str) -> list[AddedLine]:
+    """Return added text lines from a zero-context unified diff."""
+
+    path: str | None = None
+    line_number: int | None = None
+    added: list[AddedLine] = []
+
+    for raw in diff.splitlines():
+        if raw.startswith("+++ "):
+            header_path = raw[4:].strip()
+            path = header_path[2:] if header_path.startswith("b/") else header_path
+            if path == "/dev/null":
+                path = None
+            continue
+
+        match = HUNK.match(raw)
+        if match:
+            line_number = int(match.group(1))
+            continue
+
+        if path is None or line_number is None:
+            continue
+        if raw.startswith("+"):
+            added.append(AddedLine(path, line_number, raw[1:]))
+            line_number += 1
+        elif raw.startswith("-") or raw.startswith("\\ No newline"):
+            continue
+        else:
+            line_number += 1
+
+    return added
+
+
+def violations_for_diff(diff: str) -> list[Violation]:
+    violations: list[Violation] = []
+    for line in parse_added_lines(diff):
+        for label, pattern in PATH_PATTERNS:
+            if pattern.search(line.text):
+                violations.append(Violation(line.path, line.line_number, label))
+    return violations
+
+
+def render_violations(violations: list[Violation]) -> str:
+    lines = [
+        "workstation-path check failed: newly added machine-specific paths found"
+    ]
+    lines.extend(
+        f"{violation.path}:{violation.line_number}: {violation.label}"
+        for violation in violations
+    )
+    lines.append("Use an environment variable, repository-relative path, or durable URL.")
+    return "\n".join(lines)
+
+
+def git_diff(diff_range: str) -> str:
+    completed = subprocess.run(
+        [
+            "git",
+            "diff",
+            "--unified=0",
+            "--no-color",
+            "--no-ext-diff",
+            diff_range,
+            "--",
+        ],
+        check=True,
+        stdout=subprocess.PIPE,
+        text=True,
+    )
+    return completed.stdout
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument(
+        "--diff-range",
+        required=True,
+        help="Git revision range to inspect, for example BASE...HEAD",
+    )
+    args = parser.parse_args()
+
+    violations = violations_for_diff(git_diff(args.diff_range))
+    if violations:
+        print(render_violations(violations), file=sys.stderr)
+        return 1
+
+    print("workstation-path check passed")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
diff --git a/scripts/tests/test_check_workstation_paths.py b/scripts/tests/test_check_workstation_paths.py
new file mode 100644
index 0000000000..d511a376a7
--- /dev/null
+++ b/scripts/tests/test_check_workstation_paths.py
@@ -0,0 +1,74 @@
+#!/usr/bin/env python3
+
+from __future__ import annotations
+
+import importlib.util
+import sys
+import unittest
+from pathlib import Path
+
+
+SCRIPT = Path(__file__).parents[1] / "check_workstation_paths.py"
+SPEC = importlib.util.spec_from_file_location("check_workstation_paths", SCRIPT)
+assert SPEC is not None and SPEC.loader is not None
+MODULE = importlib.util.module_from_spec(SPEC)
+sys.modules[SPEC.name] = MODULE
+SPEC.loader.exec_module(MODULE)
+
+
+class WorkstationPathCheckTests(unittest.TestCase):
+    def test_added_workstation_path_is_reported_without_its_value(self) -> None:
+        private_path = "/" + "home" + "/example/project/output.log"
+        diff = "\n".join(
+            [
+                "diff --git a/report.md b/report.md",
+                "--- a/report.md",
+                "+++ b/report.md",
+                "@@ -2,0 +3 @@",
+                f'+artifact = "{private_path}"',
+            ]
+        )
+
+        violations = MODULE.violations_for_diff(diff)
+        self.assertEqual(len(violations), 1)
+        self.assertEqual(violations[0].path, "report.md")
+        self.assertEqual(violations[0].line_number, 3)
+        rendered = MODULE.render_violations(violations)
+        self.assertNotIn(private_path, rendered)
+
+    def test_removed_paths_and_portable_values_are_ignored(self) -> None:
+        private_path = "/" + "Users" + "/example/project/output.log"
+        diff = "\n".join(
+            [
+                "diff --git a/report.md b/report.md",
+                "--- a/report.md",
+                "+++ b/report.md",
+                "@@ -3 +3 @@",
+                f'-artifact = "{private_path}"',
+                '+artifact = "${ARTIFACT_ROOT}/output.log"',
+            ]
+        )
+
+        self.assertEqual(MODULE.violations_for_diff(diff), [])
+
+    def test_root_tilde_and_windows_home_paths_are_detected(self) -> None:
+        root_path = "/" + "root" + "/work/output.log"
+        tilde_path = "~" + "/work/output.log"
+        windows_path = "C:" + "\\" + "Users\\example\\output.log"
+        diff = "\n".join(
+            [
+                "diff --git a/report.md b/report.md",
+                "--- a/report.md",
+                "+++ b/report.md",
+                "@@ -0,0 +1,3 @@",
+                f"+{root_path}",
+                f"+{tilde_path}",
+                f"+{windows_path}",
+            ]
+        )
+
+        self.assertEqual(len(MODULE.violations_for_diff(diff)), 3)
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/spec/adr/ADR-006-own-quantizer.md b/spec/adr/ADR-006-own-quantizer.md
index f274b1456e..756913817c 100644
--- a/spec/adr/ADR-006-own-quantizer.md
+++ b/spec/adr/ADR-006-own-quantizer.md
@@ -24,7 +24,8 @@ The `turbo-quant` crate (v0.1) was the original plan. Investigation revealed thr
 
 ## Decision
 
-**Drop `turbo-quant` crate.** Extract the quantizer core from `~/dev_bak/TurboQuantDB/` and adapt for pgrx.
+**Drop `turbo-quant` crate.** Extract the quantizer core from
+`jyunming/TurboQuantDB` and adapt for pgrx.
 
 ### What to extract
 
diff --git a/spec/adr/ADR-024-fwht-transform-strategy.md b/spec/adr/ADR-024-fwht-transform-strategy.md
index 23164d0eb4..d871c8f970 100644
--- a/spec/adr/ADR-024-fwht-transform-strategy.md
+++ b/spec/adr/ADR-024-fwht-transform-strategy.md
@@ -43,7 +43,7 @@ Configuration A is retired. The decision is between B and C.
 
 ### What TurboQuantDB Does Differently
 
-The upstream TurboQuantDB implementation (`~/dev_bak/TurboQuantDB/`) uses configuration C:
+The upstream `jyunming/TurboQuantDB` implementation uses configuration C:
 
 1. **Pad to `n = next_power_of_two(d)`** — for d=1536, n=2048
 2. **Full FWHT on n dimensions** — complete decorrelation across all coordinates
@@ -156,4 +156,5 @@ No code path should produce configuration A going forward.
 - Review 202: A4 1536 tail-truncation probes — transform-tail truncation identified as dominant loss
 - Review 203: A4 1536 tiled-FWHT quantizer — production tiled FWHT implementation
 - Review 204: A4 full vs tiled FWHT — TurboQuantDB comparison and three-configuration analysis
-- TurboQuantDB source: `~/dev_bak/TurboQuantDB/src/quantizer/prod.rs`, `mse.rs`, `../linalg/hadamard.rs`
+- TurboQuantDB source: `src/quantizer/prod.rs`, `mse.rs`, and
+  `src/linalg/hadamard.rs` in `jyunming/TurboQuantDB`
diff --git a/spec/adr/ADR-045-symphonyqg-quantized-graph-access-method.md b/spec/adr/ADR-045-symphonyqg-quantized-graph-access-method.md
index 1136952ff9..735aef6d1a 100644
--- a/spec/adr/ADR-045-symphonyqg-quantized-graph-access-method.md
+++ b/spec/adr/ADR-045-symphonyqg-quantized-graph-access-method.md
@@ -257,8 +257,7 @@ added a pragmatic q ∈ {1, 2, 4, 8} extension to `RaBitQQuantizer`
 so non-Symphony consumers (DiskANN in-memory tier, general
 prefilter, offline eval) can dial the recall/storage tradeoff.
 The slice-12 q-bit implementation is not paper-faithful to
-**Extended RaBitQ** (Gao et al., SIGMOD 2025, arXiv:2409.09913 —
-`~/dev_bak/papers/extended-rabitq-2025-sigmod-arxiv-2409.09913.pdf`):
+**Extended RaBitQ** (Gao et al., SIGMOD 2025, arXiv:2409.09913):
 
 1. **Scalar quantizer.** Slice 12 uses uniform binning on
    `±2σ`; Extended RaBitQ uses a Lloyd-Max-like codebook
diff --git a/spec/adr/ADR-085-ec-distann-single-global-graph.md b/spec/adr/ADR-085-ec-distann-single-global-graph.md
index 9c8d783839..76ab444abb 100644
--- a/spec/adr/ADR-085-ec-distann-single-global-graph.md
+++ b/spec/adr/ADR-085-ec-distann-single-global-graph.md
@@ -37,8 +37,7 @@ branch + packet path):
 
 The root cause is architectural: a lossy partition-level routing decision
 must be hedged wider as recall targets rise and corpora grow. DistributedANN
-(**arXiv:2509.06046**, the durable citation; convenience local copy
-`~/dev_bak/papers/distributedann-2509.06046.pdf`; same research group as
+(**arXiv:2509.06046**, the durable citation; same research group as
 SPANN/SPIRE, replaced partition-routing in Bing production) inverts the
 design: distribute the storage, keep the index whole. One global Vamana
 graph; records hash-placed; query cost = beam × hops, corpus-independent.
diff --git a/spec/reviews/evidence.md b/spec/reviews/evidence.md
index afd1a21a23..73e1954a46 100644
--- a/spec/reviews/evidence.md
+++ b/spec/reviews/evidence.md
@@ -57,4 +57,4 @@ commit" convention.
 | FND-009 | low | NFR-017's "p50 latency at matched recall" does not define the operating-point selection rule (interpolation across sweep points vs nearest cell at ≥ 0.999), and the p50/p95 rows are specified only at 100k while the recall row spans 10k/50k/100k — pre-register the matched-recall protocol (the Task 146 packet's rule, if that is the intent) in the Verification section so the four-way table is not assembled ad hoc | NFR-017, TC-044 |
 | FND-010 | low | NFR-020's fault taxonomy includes "network partition", but the gate substrate is loopback multi-instance (ADR-085 D2) where a true partition is not injectable with the reused connection-level drill machinery (reset/timeout/termination are); name the injection mechanism for the partition case or scope it to connection-level manifestations so the "100% drill matrix" row is executable | NFR-020, TC-042 |
 | FND-011 | low | FR-077-CON-4 / ADR-085 D8 bound stitch memory by "one vec_id group plus prune working set" with validation "Analysis + build instrumentation", but no numeric bound and no named analysis artifact path — non-test verification methods must name their artifact (analysis doc + which epoch-manifest/instrumentation field). As written the CON is not falsifiable | FR-077, ADR-085 |
-| FND-012 | low | ADR-085 cites the DistributedANN paper by a local home-directory path (`~/dev_bak/papers/distributedann-2509.06046.pdf`) — a non-durable, single-machine reference; the arXiv id (2509.06046) alone is the durable citation. Minor because it is background literature, not gate evidence, but the packet-discipline rule against citing local-only paths applies to ADRs too | ADR-085 |
+| FND-012 | resolved | ADR-085 now cites DistributedANN by its durable arXiv identifier rather than a local home-directory path. | ADR-085; arXiv:2509.06046 |

From a21e844bf05ddd68a1ca895be3d08fa11a213080 Mon Sep 17 00:00:00 2001
From: Agent IX <agent@agent-ix.dev>
Date: Sat, 29 Aug 2026 02:52:47 -0700
Subject: [PATCH 2/2] docs: request Task 240 path hygiene review

---
 plan/tasks/240-artifact-path-sanitization.md  |  5 +-
 plan/tasks/README.md                          |  2 +-
 .../artifacts/manifest.md                     | 57 +++++++++++++
 .../artifacts/redaction-audit.md              | 37 +++++++++
 .../001-current-tree-sanitization/request.md  | 82 +++++++++++++++++++
 5 files changed, 180 insertions(+), 3 deletions(-)
 create mode 100644 reviews/task-240/001-current-tree-sanitization/artifacts/manifest.md
 create mode 100644 reviews/task-240/001-current-tree-sanitization/artifacts/redaction-audit.md
 create mode 100644 reviews/task-240/001-current-tree-sanitization/request.md

diff --git a/plan/tasks/240-artifact-path-sanitization.md b/plan/tasks/240-artifact-path-sanitization.md
index 61d929f674..99cd853425 100644
--- a/plan/tasks/240-artifact-path-sanitization.md
+++ b/plan/tasks/240-artifact-path-sanitization.md
@@ -1,7 +1,8 @@
 # Task 240: Artifact Path Sanitization
 
-Status: **implementation in progress** (2026-08-29). Priority: P0 privacy and
-evidence hygiene. Origin: private Engineering Assurance gap audit.
+Status: **implementation complete; packet 001 review-open; history remediation
+pending explicit decision** (2026-08-29). Priority: P0 privacy and evidence
+hygiene. Origin: private Engineering Assurance gap audit.
 
 ## Why
 
diff --git a/plan/tasks/README.md b/plan/tasks/README.md
index 8e7ae62171..a89928ade9 100644
--- a/plan/tasks/README.md
+++ b/plan/tasks/README.md
@@ -279,7 +279,7 @@ These pre-lane task files are retained only for historical context under
 238. `238-ec-distann-retry-snapshot-use-after-free.md` — **implementation and evidence merged; packet 001 outside closeout review-open** (updated 2026-08-27, P0 correctness): the equivalent `RegisteredSnapshotGuard` lifetime fix remains on main via `15f7fcf5f` / Task 167 PR #77. Checkpoint `7d4103885` restores the deterministic forced-retry regression and explicit loopback-only plaintext test conninfo; both PG18 callers pass on base `3c81319a3`. PR #91 merged the implementation and evidence as `bcc9b0bfa`; only the outside ACCEPT remains open. Evidence and request: `reviews/task-238/001-retry-snapshot-uaf/`.
 239. `239-ec-distann-bounded-read-overfetch.md` — **complete — review-closed ACCEPT; HARNESS REGRESSION CORRECTED, EXACT-MAIN LAZY-10 SEMANTIC PATH RESTORED TO 10/10; no rerun; packet 004 10k/50k/100k NOT triggered; Task 229 / 230--233 semantic-surface blocker LIFTED** (updated 2026-08-26, P1 correctness/performance): packet 001's 12/10 result was a shared-session batch-size GUC leak in the benchmark harness — a variant at batch size 10 emitted no `SET`, so the "lazy-10" arm inherited the eager control's 0 — not production overfetch, and the bound was never widened. Packet 002 corrected the harness but its cross-SHA runner aborted pre-semantics on 40-versus-37 stage rows. Packet 003 removed that confounder with exact-main runner/extension plus only the harness fix; its sole C1--C5 run passed every preregistered gate — nine scenarios exactly once in both logs, seven core rows at `control_batch_size=0 candidate_batch_size=10` with exact identity and zero duplicates, `exactly_one_window` 6 remote + 4 local = 10 reads/bound 10, mixed and outage pass, routed DELETE+VACUUM pass, both recall arms 0.9990 over 200 queries / 2,000 trials with predictions byte-identical to packet 001 — and the stopped 1.2 GB fixture was removed after capture. `git diff 41392c011 def565270 -- src` is empty, so no production runtime behavior changed and the 10k/50k/100k closeout matrix does not apply. Determinism rests on the same 6/4/10 split and digests at three independent extension SHAs (task-191 `7883cfcf`, task-198 `2ff72b3e`, this run `4ab2aa9a9`), so no rerun was needed. Carried follow-ups for whichever task next touches the multinode semantic harness: mixed/outage failures emit no structured `pass=false` row; `owner_payload_plan_cache` retains the same conditional-`SET` leak shape; recall/latency child batch size stays inferred. Closeout verdict: `reviews/task-239/003-main-baseline-semantic-proof/feedback/2026-08-26-02-reviewer.md`; result decision: `reviews/task-239/003-main-baseline-semantic-proof/artifacts/live-run-decision.md`; review request: `reviews/task-239/003-main-baseline-semantic-proof/request.md`; live authorization: `reviews/task-239/003-main-baseline-semantic-proof/feedback/2026-08-26-01-reviewer.md`.
 
-240. `240-artifact-path-sanitization.md` — **implementation in progress** (2026-08-29, P0 privacy/evidence hygiene): remove current-tree internal scratch locators, require portable comparison-repository inputs, and add a non-echoing diff guard for newly introduced workstation paths; historical Git remediation remains a separate explicit decision.
+240. `240-artifact-path-sanitization.md` — **implementation complete; packet 001 review-open; history remediation pending explicit decision** (2026-08-29, P0 privacy/evidence hygiene): current-tree internal scratch locators are removed, comparison-repository inputs are explicit, and a non-echoing diff guard rejects newly introduced workstation paths; packet `reviews/task-240/001-current-tree-sanitization/` requests outside review, while historical Git remediation remains a separate explicit decision.
 
 ## Coordination rules
 
diff --git a/reviews/task-240/001-current-tree-sanitization/artifacts/manifest.md b/reviews/task-240/001-current-tree-sanitization/artifacts/manifest.md
new file mode 100644
index 0000000000..4d0317b9b1
--- /dev/null
+++ b/reviews/task-240/001-current-tree-sanitization/artifacts/manifest.md
@@ -0,0 +1,57 @@
+# Task 240 packet 001 artifact manifest
+
+- Task/packet: `task-240/001-current-tree-sanitization`
+- Baseline: `23fb9b7ba1f0803be5dfc700d9865f80fbf60862`
+- Implementation: `e393515c7c94e0f127bcbac633052ee457c0527e`
+- Review-request checkpoint: this packet commit
+- Scope: current tracked tree and future added diff lines
+- Excluded: Git-history rewrite, hosted hidden refs/caches, and bulk
+  normalization of unrelated legacy workstation paths
+
+## Audit results
+
+- Baseline files containing the identified internal scratch locator: 31.
+- Implementation files containing that locator: 0.
+- Baseline files containing the audited absolute workstation prefix: 6,396.
+- Implementation files containing that exact prefix: 6,390.
+- Added-lines guard across baseline to implementation: pass.
+- Whitespace/diff validation: pass.
+
+The sensitive locator and matching source text are intentionally absent from
+this public packet. The audit reports counts and categories only.
+
+## Functional validation
+
+- `python3 scripts/tests/test_check_workstation_paths.py`: 3 tests pass.
+- `python3 scripts/check_workstation_paths.py --diff-range
+  23fb9b7ba1f0803be5dfc700d9865f80fbf60862...e393515c7c94e0f127bcbac633052ee457c0527e`:
+  pass.
+- `cargo fmt --all -- --check`: pass with existing stable-rustfmt warnings.
+- Isolated `cargo build -p ecaz-cli`: pass with one existing dead-code
+  warning for `LoadedDistributedPlacementConfig.path`.
+- `cargo test -p ecaz-cli commands::dev::install`: compile pass; 546 tests
+  filtered out and zero selected.
+- Freshly built CLI with each repository variable unset:
+  - `ecaz dev install pgvector --pg 18`: exit 2; `--repo` required.
+  - `ecaz dev install vectorscale --pg 18`: exit 2; `--repo` required.
+
+No installer runtime path was reached during the missing-argument checks.
+
+## Known baseline test-runner failures
+
+`scripts/tests/run.sh` reports:
+
+- three failures in
+  `scripts/tests/test_pg17_scratch_psql_socket_resolution.py`;
+- four errors in
+  `scripts/tests/test_resolve_scratch_socket_dir.py`.
+
+Both groups arise because the unchanged baseline references already-deleted
+scripts. They are outside Task 240 and are recorded without being reclassified
+as a Task 240 regression.
+
+## Generated artifact cleanup
+
+The isolated Cargo target directory used for the behavioral CLI checks was
+removed after validation. It contained generated build output only; no tracked
+file or installed extension was removed.
diff --git a/reviews/task-240/001-current-tree-sanitization/artifacts/redaction-audit.md b/reviews/task-240/001-current-tree-sanitization/artifacts/redaction-audit.md
new file mode 100644
index 0000000000..32b8faafc1
--- /dev/null
+++ b/reviews/task-240/001-current-tree-sanitization/artifacts/redaction-audit.md
@@ -0,0 +1,37 @@
+# Task 240 current-tree redaction audit
+
+## Classification
+
+The 31 baseline files containing the identified internal scratch locator were:
+
+| Category | Files | Treatment |
+|---|---:|---|
+| CLI source and documentation | 2 | Removed implicit local defaults; documented explicit operator inputs |
+| Benchmark/reviewer evidence | 2 | Replaced only source/build locator text |
+| Planning and design | 4 | Replaced local locations with durable public identities |
+| Review packets | 18 | Replaced citation/source locators; retained findings and measurements |
+| Specifications, ADRs, and findings | 5 | Replaced local citations and closed the resolved citation finding |
+| **Total** | **31** | **Current-tree locator count is zero** |
+
+## Evidence-integrity check
+
+The diff was reviewed line by line. Existing evidence edits are confined to:
+
+- paper-location to arXiv-identifier substitutions;
+- checkout-location to public repository/source-identity substitutions;
+- local build-path components to a neutral `[local-checkout]` marker; and
+- prose wrapping needed by those substitutions.
+
+No measured value, result row, duration, source SHA, algorithm description,
+review verdict, or acceptance conclusion changed. The edited historical build
+log is described by its manifest but is not content-digest-bound there.
+
+## Residual exposure
+
+This is a current-tree repair, not a history purge. Existing Git objects retain
+the old bytes, and the legacy corpus still contains other workstation paths.
+The new guard prevents newly added instances; it does not retroactively reject
+or silently rewrite pre-control evidence.
+
+Any later history remediation must have its own authorization, recovery plan,
+force-push coordination, hidden-ref/cache verification, and post-rewrite audit.
diff --git a/reviews/task-240/001-current-tree-sanitization/request.md b/reviews/task-240/001-current-tree-sanitization/request.md
new file mode 100644
index 0000000000..94198e7d93
--- /dev/null
+++ b/reviews/task-240/001-current-tree-sanitization/request.md
@@ -0,0 +1,82 @@
+---
+task: 240
+packet: 001-current-tree-sanitization
+agent: Codex
+role: coder
+model: gpt-5
+date: 2026-08-29
+seq: 01
+---
+
+# Task 240 current-tree sanitization review request
+
+Please review the current-tree privacy repair and future-publication guard at
+`e393515c7c94e0f127bcbac633052ee457c0527e`, based on public main
+`23fb9b7ba1f0803be5dfc700d9865f80fbf60862`.
+
+## Requested decision
+
+Accept the implementation if the reviewer agrees that:
+
+1. the 31 current-tree references to the identified internal scratch locator
+   were replaced with durable public citations, source identities, explicit
+   operator inputs, or a neutral local-checkout marker;
+2. those edits changed locator/citation metadata only and did not alter
+   benchmark measurements, source SHAs, findings, or technical conclusions;
+3. the CLI must require `--repo` or the documented environment variable for
+   both external comparison-extension installers;
+4. the added-lines guard safely reports file, line, and category without
+   echoing the matched value; and
+5. Git-history remediation remains explicitly out of scope until a separate
+   destructive-history decision is approved.
+
+## Implementation
+
+- Removed the identified internal scratch locator from every tracked file in
+  the current tree.
+- Replaced paper locations with arXiv identifiers and source checkouts with
+  durable repository identities plus pinned SHAs where already available.
+- Removed machine-specific comparison-repository defaults from
+  `ecaz dev install pgvector` and `ecaz dev install vectorscale`.
+- Added `scripts/check_workstation_paths.py`, focused synthetic tests, and a
+  lightweight pull-request/main-push workflow.
+- Added the workstation-path rule to `AGENTS.md`.
+
+The guard covers added Unix, root, Windows, and tilde-home paths. It inspects
+only added diff lines so pre-existing immutable evidence does not cause every
+unrelated pull request to fail.
+
+## Deliberate boundary
+
+The baseline contains 6,396 tracked files with one common absolute workstation
+prefix. This checkpoint reduces that exact-prefix count to 6,390 but does not
+bulk-rewrite the legacy evidence corpus. It does reduce the specifically
+identified internal scratch locator from 31 tracked files to zero.
+
+Old values remain reachable from existing Git objects. No branch, tag, pull
+request ref, commit, or repository history was rewritten or deleted. A
+history-wide purge would require a coordinated force-push, collaborator
+re-clone procedure, and hosting-provider cache/ref handling; that is not
+authorized by this packet.
+
+## Validation
+
+- Focused Python tests: 3/3 pass.
+- Added-lines guard over `23fb9b7...e393515c7`: pass.
+- Current-tree exact-locator audit: zero matches.
+- `git diff --check`: pass.
+- `cargo fmt --all -- --check`: pass (existing stable-rustfmt warnings only).
+- Isolated `cargo build -p ecaz-cli`: pass (one existing unread-field warning).
+- `cargo test -p ecaz-cli commands::dev::install`: compile pass; the filter
+  selected zero tests, so this is not claimed as behavioral coverage.
+- New binary, repository variables unset:
+  - pgvector installer: exit 2, missing required `--repo <REPO>`;
+  - vectorscale installer: exit 2, missing required `--repo <REPO>`.
+
+The repository-wide `scripts/tests/run.sh` is not green on the unchanged
+baseline: it reports three failures and four errors in the socket-resolution
+tests because two referenced shell scripts are already absent. This task does
+not modify that unrelated surface and does not claim the full runner as a pass.
+
+See `artifacts/manifest.md` and `artifacts/redaction-audit.md` for the exact
+scope and evidence classification.