LR's 'unknown unit_id' skip label and counter describe out-of-batch ids as unknown — 113 of 210 campaign skips were real dataset units (likely cosmetic: no promotion was lost; 18/18 promotable skips were already reachable)

[gadievron]

Summary (likely cosmetic)

The llm-reachability skip line and counter describe out-of-batch ids as
unknown: signal #N: unknown unit_id '<id>' … — skipped in batch i/N
and signals_skipped_unknown_unit. But "unknown" here means only "not in
the current sub-batch" — the id is often a perfectly real unit of the same
scan, one batch over. On the campaign corpus, 113 of 210 skips (54%) were
ids that exist in the scan's own dataset. No promotion was lost: all 18
promotable (entry_point/high) skipped units were already reachable or entry
points, so the corpus impact is confined to the misleading label and
counter. The drop itself is defensible — the model never saw that unit's
code in this batch — this issue is about the label's semantics, not an ask
to admit out-of-batch signals.

Mechanism (at ad2bb7e)

1. core/llm_reachability.py:362-371 — a signal whose unit_id is not in
   valid_unit_ids logs the "unknown unit_id" line and fires
   on_signal_skip(reason="unknown_unit_id", …).
2. The caller passes valid_unit_ids=batch_ids where
   batch_ids = {u["id"] for u in sub_batch} (:784, :811-812) — the
   CURRENT SUB-BATCH, not the scan's unit set. The prompt projection sends
   only unit_id+code per unit (:171-189), so an out-of-batch id is
   one the model inferred from a call site, not one it was shown — the skip
   is batch-scoped by design of the batching, and the label says "unknown",
   which reads as "not a real unit".
3. Lineage: closed LLM-reach report gaps: the unknown-unit_id skip line loses kind/confidence and has no counter; entry_points_promoted is not decomposed #602 added the kind/confidence/batch label and the
   counter; closed fix(llm-reach): resume parity — per-unit records under the backend-identity gate (#532) #543's disclosed behavior change says "cross-batch ids
   were hallucination-only inputs" — the corpus partition below refutes
   that premise (113/210 were real dataset units).

Executed at ad2bb7e

$ python3 up19_probe.py
signal for a real out-of-batch unit id, kind=entry_point, confidence=high:
  log line: ["signal #0: unknown unit_id 'real.c:helper' (kind='entry_point', confidence='high') — skipped in batch 1/2"]
  skip reason recorded: ['unknown_unit_id']
  the id exists in the scan's dataset: True
returned signals: 0

up19_probe.py — full script

#!/usr/bin/env python3
"""UP19: the LR skip line labels an out-of-batch REAL unit id as 'unknown'.
Run from the OpenAnt checkout root."""
import sys, os, json, tempfile
assert os.path.isfile("libs/openant-core/pyproject.toml"), "run from the repo root"
sys.path.insert(0, "libs/openant-core")
from core.llm_reachability import parse_response

scan_units = ["app.c:main", "real.c:helper", "util.c:f"]        # the scan's dataset
batch_ids  = {"app.c:main", "util.c:f"}                          # the current sub-batch
reply = json.dumps({"signals": [
    {"unit_id": "real.c:helper", "kind": "entry_point", "confidence": "high",
     "reason": "called from main"}]})

log_lines, skips = [], []
sigs = parse_response(reply, valid_unit_ids=batch_ids, on_error=log_lines.append,
                      on_signal_skip=lambda **kw: skips.append(kw["reason"]),
                      batch_label="batch 1/2")
print("signal for a real out-of-batch unit id, kind=entry_point, confidence=high:")
print("  log line:", log_lines)
print("  skip reason recorded:", skips)
print("  the id exists in the scan's dataset:", "real.c:helper" in scan_units)
print("returned signals:", len(sigs))

Campaign receipts (committed run artifacts, not reader-runnable): 210 skips
across the 25 LR-bearing runs — 113 ids exist in the run's own
dataset.json, 97 do not; per-run examples include
fstyp.tproj/fstyp_ntfs.c:main (apple-diskdev) and
notifyd/notifyd.c:main (libnotify, skipped in batches 8/12 and 9/12 —
both real). All 18 promotable (entry_point/high) real skips were ANALYZED
in their runs' results.json with reachable=True (17/18 already
is_entry_point=True) — the negative result this filing's "likely
cosmetic" label rests on.

Falsifier

A promotable (entry_point/high) out-of-batch skip whose unit was NOT
already reachable and ended in pruned_units.json — that upgrades this to
a recall defect. None in 25 runs. Authored at filing time; no falsifier
existed in the source record.

Prior art

Closed #602 (the label/counter's own issue — delivered); closed #543 (the
resume-parity change whose "hallucination-only" premise the partition
corrects). No open issue covers the label semantics.

Fix-direction: neither — observability wording: the line and counter name
what was skipped accurately (e.g. "out-of-batch unit_id"); no finding or
skip decision changes.