fix(search): hashes and commit trailers are not identifiers

AGENT_DIARY.md

@@ -57,6 +57,9 @@
   Пункты «RU→EN словарь» и «concept→symbol словарь» сняты: их заменяет контракт.
 - **Урок:** не учить сервер языкам, которые знает хост; отдавать `hints` и честно
   просить их, когда в вопросе нет code-имени.
+- **E71 (Change-Id):** хеш из Gerrit занимал первый слот идентификаторов, гейт считал
+  `0` вхождений и отказывал на каждом коммите sqlalchemy (0/40). Фильтр хешей и
+  трейлеров поднял tool до 11/40, B1 dev не изменился (19/38/22/2).
 - **E70 (внешний репо):** на `psf/requests` grep берёт 19/25 hit@5, search 15,
   neighbours 13, наш tool 10. На маленьком корпусе оркестрация проигрывает обычному
   поиску по тексту; «мы лучше grep» верно только для символьного поиска на нашей панели.

EXPERIMENTS_LOG.md

@@ -1,4 +1,28 @@
 # EXPERIMENTS_LOG.md — Audit Verification (2026-07-22)
+## [2026-11-10] — E71: a Gerrit `Change-Id` made the gate abstain on every commit
+
+**Found by** the sqlalchemy run of B2: tool 0/40 at hit@5 while neighbours got
+18. Not a ranking result - `_identifiers` took the 40-char `Change-Id` hash as
+the first identifier, the gate counted `0` occurrences, and the answer was
+`abstain` on every commit that carried one.
+
+**Fix:** hashes (32+ hex, with or without the Gerrit `I` prefix) and commit
+trailers (`fixed`, `fixes`, `issue`, `closes`, `resolves`, `changelog`) are no
+longer identifiers; `close` stays, it is a real method name.
+
+```
+sqlalchemy, 40 commits   hit@5 before -> after
+tool                          0  ->  11
+neighbours                   18  ->  19
+search                       18  ->  18
+grep                         20  ->  20
+B1 dev (71)              19/38/22/2 -> identical
+```
+
+After the fix the tool is still below grep on this external repo (11 vs 20), so
+the small/medium-repo conclusion stands; what changed is that the abstain is no
+longer systematic.
+
 ## [2026-11-10] — E70: external repo (psf/requests) — plain grep wins
 
 **Why:** B1 still lives in our own repository. Here the repo, history, messages

experiments/bench_2026_11/REPORT.md

@@ -141,8 +141,40 @@ visible in the log before the crash; no code, arm or metric changed after that
 (the arms were frozen in Amendments 1-3 before the sealed run), so this is
 disclosed rather than "clean".
 
-## B2 — external repository (`psf/requests`)
+## B2 — external repository: `sqlalchemy` (the finding that mattered)
+
+679 `.py` files, corpus at `8b62343` (2026-09-24), 34 984 chunks, one index
+build (53 min). 40 commits.
+
+```
+                 hit@1  hit@5         hit@1  hit@5
+                 before E71                 after E71
+tool                 0      0              8     11
+neighbours          13     18             11     19
+search              12     18             12     18
+grep                 7     20              7     20
+```
+
+**The zero was a bug, not a small-repo effect.** `_identifiers` took the Gerrit
+`Change-Id` (a 40-char hash) as the first identifier, the gate counted
+`Iaf8be… = 0` occurrences, and every such commit abstained:
+
+```
+before: identifiers: Iaf8be4a9…, _ConnectionRecord.__close, run_scenario, Fixed
+        verdict: abstain (0.00)
+after:  identifiers: _ConnectionRecord.__close, run_scenario, AnyIO, sqlalchemy.testing.cancellation
+        verdict: answer (0.75), gold at row 2
+```
+
+E71 drops hashes (32+ hex, with or without the Gerrit `I` prefix) and commit
+trailers (`fixed`, `fixes`, `issue`, `closes`, …) from `_identifiers`; `close`
+stays, because it is a real method name. B1 dev is **identical** after the fix
+(19 / 38 / 22 / 2), so the calibrated corpus did not move.
 
+After the fix the tool is still below grep here (11 vs 20) — the small-repo
+conclusion from `requests` survives — but the catastrophic abstain is gone.
+
+## B2 — external repository (`psf/requests`)
 Nothing in this run is ours: the repository, its history, its commit messages and
 the gold all come from other people. Same rule as B1; corpus at
 `e331a288` (2026-03-24), 64 files / 786 chunks, one index build (16 s).
@@ -184,6 +216,24 @@ split was re-run on the merged build: **identical** (`tool` 19, `neighbours` 38,
 `search` 22, `grep` 2; paired +16). The fix does not touch this benchmark, which
 is what a classification change should look like.
 
+## Reproduce
+
+```
+# B1 corpus (this repo at the frozen base)
+git worktree add D:\Project\bench-B 7c2c7a3c6417539bc08f807d6790fe6e722d3a84
+python scripts/reindex_text_only.py --project D:\Project\bench-B
+python experiments/bench_2026_11/commit_tasks.py
+python experiments/bench_2026_11/run.py --split dev
+python experiments/bench_2026_11/run.py --split sealed --sealed   # once per release
+
+# B2 corpus (external)
+git clone https://github.com/psf/requests D:\Project\bench-ext\requests
+git -C D:\Project\bench-ext\requests worktree add D:\Project\bench-ext\req-B e331a288
+python scripts/reindex_text_only.py --project D:\Project\bench-ext\req-B
+python experiments/bench_2026_11/external.py --repo D:\Project\bench-ext\requests \
+    --corpus D:\Project\bench-ext\req-B --base e331a288 --limit 40
+```
+
 ## Threats
 
 - The corpus is one snapshot at `B`. A file renamed after `B` makes its task

experiments/bench_2026_11/results_external_sqlalchemy.json

@@ -0,0 +1,24 @@
+{
+ "repo": "sqlalchemy",
+ "base": "8b62343934387efcc68f02032fa6cc07ac65ba26",
+ "tasks": 40,
+ "table": {
+  "tool": {
+   "top1": 8,
+   "top5": 11
+  },
+  "neighbours": {
+   "top1": 11,
+   "top5": 19
+  },
+  "search": {
+   "top1": 12,
+   "top5": 18
+  },
+  "grep": {
+   "top1": 7,
+   "top5": 20
+  }
+ },
+ "invalid": 0
+}

src/mcp/tools/investigate_tool.py

@@ -51,13 +51,21 @@
 _MAX_IDS = 4
 _CHANNEL_TIMEOUT = 12.0
 _GREP_TIMEOUT = 15.0
+# Hashes are not identifiers: a Gerrit `Change-Id` (40 hex chars) or a commit
+# sha can never occur in the source, and counting them made the gate abstain on
+# every sqlalchemy commit (E71, external benchmark: 0/40 at hit@5).
+_HEX_TOKEN_RE = re.compile(r"^I?[0-9a-fA-F]{32,}$")
 _STOPWORDS = {
     "the", "and", "for", "how", "why", "what", "does", "used", "use", "where",
     "when", "which", "with", "from", "into", "that", "this", "files", "file",
     "changed", "change", "checker", "calls", "call", "work", "works", "after",
     "who", "whom", "whose", "defined", "definition", "define", "defines",
     "implemented", "implement", "implements", "located", "sits", "lives",
     "raises", "raise", "returns", "return", "callers", "caller", "usages",
+    # Commit trailers (Gerrit/GitHub) are not code names; counting them made the
+    # gate abstain on prose (E71). `close` is left out on purpose: it is a real
+    # method name.
+    "fixed", "fixes", "issue", "issues", "resolves", "closes", "changelog",
 }
 
 
@@ -90,6 +98,8 @@ def _identifiers(query: str) -> List[str]:
         token = token.strip(".")
         if not token or token.lower() in _STOPWORDS:
             continue
+        if _HEX_TOKEN_RE.match(token):
+            continue
         if token not in found:
             found.append(token)
 

tests/test_identifiers.py

@@ -0,0 +1,46 @@
+"""Identifiers extracted from a question must not be hashes or commit trailers.
+
+E71, discovered by the external benchmark: a Gerrit `Change-Id` took the first
+slot in `_identifiers`, the gate counted it as "0 occurrences", and every
+sqlalchemy commit abstained (0/40 at hit@5).
+"""
+
+from __future__ import annotations
+
+from src.mcp.tools.investigate_tool import _identifiers
+
+GERRIT_MESSAGE = (
+    "discard pooled connection when its close is interrupted\n\n"
+    "Fixed issue where a task was cancelled while closing.\n\n"
+    "Fixes: #13673\n"
+    "Change-Id: Iaf8be4a9f4cc0f4435ae97ef8d961ce89375ffcc\n"
+)
+
+
+def test_gerrit_change_id_is_not_an_identifier():
+    ids = _identifiers(GERRIT_MESSAGE)
+
+    assert not any(len(i) >= 32 for i in ids), ids
+    assert "Iaf8be4a9f4cc0f4435ae97ef8d961ce89375ffcc" not in ids
+
+
+def test_commit_trailers_are_not_identifiers():
+    ids = [i.lower() for i in _identifiers(GERRIT_MESSAGE)]
+
+    for word in ("fixed", "fixes", "issue", "issues", "closes", "resolves"):
+        assert word not in ids, ids
+
+
+def test_a_real_method_name_still_survives():
+    ids = _identifiers("where is _ConnectionRecord.__close defined")
+
+    assert ids == ["_ConnectionRecord.__close"]
+
+
+def test_close_is_still_an_identifier():
+    # Left out of the stopword list on purpose: `close` is a real method name.
+    assert "close" in _identifiers("where is close defined")
+
+
+def test_a_plain_question_still_yields_its_identifier():
+    assert _identifiers("where is get_progress_file defined") == ["get_progress_file"]