Prevent workstation path publication

.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"

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

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

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.

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
 ```
 

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")
 }

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
 

plan/tasks/240-artifact-path-sanitization.md

@@ -0,0 +1,43 @@
+# Task 240: Artifact Path Sanitization
+
+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
+
+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.

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

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…"*

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
 

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 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
 
 - Freeze binary datum layout before downstream work expands.

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
 

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
 

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
 

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
-

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).