fix(store): close 3 aggregate-query tiebreaker sites flagged in #36 §3

[yakuphanycl]

§0 TL;DR

Closes the 3 aggregate-query tiebreaker sites that PR #36 §3 reported but did not silently fix per the "report, don't silently fix" hard rule from #33. Same SQLite tiebreaker bug class as #30 / #33 / #34 / #36 but DIFFERENT fix shape — these are GROUP BY queries where the deterministic tiebreaker is the GROUP key (alphabetical ASC), not rowid.

Authored by Codex (OpenAI), PR opened by Claude A — Codex's PR creation was blocked by the same `gh` 401 / GitHub connector 403 auth issue we hit 5 times already today. 6-of-6 rescue today per `feedback_codex_pr_creation_blocked.md`. Branch pushed cleanly; PR-creation rescue only.

+3 tests, 131 → 134 passing. After this PR, the SQLite-tiebreaker class is FULLY closed in the instinct repo (per-row + LIMIT-1 + aggregate, all sites covered).

§1 Sites fixed (all src/instinct/store.py)

Function	Line	Pre-fix ORDER BY	Group key	Fix
`trending()` real-delta path	~635	`... GROUP BY pattern ORDER BY observations DESC LIMIT ?`	`pattern`	append `, pattern ASC`
`effectiveness()`	~735	`... GROUP BY pattern ORDER BY confirmed DESC, suggested DESC`	`pattern`	append `, pattern ASC`
`stats()` by_category	~829	`... GROUP BY category ORDER BY count DESC`	`category`	append `, category ASC`

§2 Why ASC on group key (different from #34/#36)

PR	Mental model	Tiebreaker	Why
#30, #33, #34, #36	Per-row queries on `instincts` / `confidence_log` / `suggest_log`	`rowid DESC` (or `id DESC` where integer PK)	Per-row queries return individual records; insertion-order surrogate is the natural deterministic tiebreaker
This PR	Aggregate (GROUP BY) queries	` ASC`	GROUP BY collapses rows into aggregate values; the only stable tiebreaker is the GROUP key itself, alphabetically ASC

ASC chosen over DESC because aggregate display benefits from alphabetical-stable order on ties (predictable, scannable). DESC would also work but ASC matches the natural "first alphabetically when values tie" convention.

§3 Tests added (3 new)

Each test creates ≥2 groups producing identical aggregate values (tie scenario), calls the function twice, asserts same order both calls AND that the order matches the alphabetical group-key tiebreaker.

• `test_trending_real_delta_deterministic_under_observations_collision`
• `test_effectiveness_deterministic_under_confirmed_suggested_collision`
• `test_stats_by_category_deterministic_under_count_collision`

§4 Validation

```
$ PYTHONPATH=src python -m pytest -q
134 passed in [time] (was 131 pre-fix; +3 new)

$ PYTHONPATH=src ruff check src/instinct/store.py tests/test_store.py
All checks passed!
```

Baseline note: task brief expected 113→116, but current master already includes #37 (PR merged after the brief was drafted), so this clone verifies the live 131→134 delta. +3 strict increase confirmed.

§5 No 4th aggregate site found (decision gate clean)

Audit during this PR: `_promote_universal()` has `GROUP BY cl.pattern` but no aggregate `ORDER BY` tie to stabilize. No 4th aggregate site discovered. Scope stays at exactly the 3 sites #36 §3 flagged.

No other decision-gate triggers. No surprises.

§6 Maintainer checklist

• Confirm group-key ASC choice (vs. DESC) for aggregate queries — see §2
• Confirm scope: only the 3 §3-flagged aggregate sites
• Note: this PR completes the SQLite-tiebreaker class closure for instinct (5 PRs total: fix(store): repair the 2 failing detect_chains tests #30, fix(store): history() tiebreaker + idx_conflog_ts direction match (#30 follow-up) #33, fix(store): close 3 SQLite tiebreaker drift sites flagged in #33 #34, fix(store): close 6 bug-class + 1 LIMIT-1 tiebreaker sites flagged in #34 §3 #36, this)

§7 Hard rule self-quote

• No `--no-verify`
• File scope: `src/instinct/store.py` + `tests/test_store.py` only
• Did NOT touch: per-row sites (fix(store): close 3 SQLite tiebreaker drift sites flagged in #33 #34/fix(store): close 6 bug-class + 1 LIMIT-1 tiebreaker sites flagged in #34 §3 #36 territory, fully closed), `src/instinct/server.py`, `tests/mcp/*` (test(mcp): add priority smoke coverage for ok envelopes (#26 §6) #35/test(mcp): extend smoke coverage to remaining 18 tools (#26 §6 follow-up) #37 territory), SKILL.md, .cursor/rules/, README, CHANGELOG, pyproject
• All 3 chartered aggregate sites fixed in ONE commit (same aggregate-tiebreaker mental model)
• Test count strictly increases: 131 → 134
• No 4th site silently extended scope (decision gate clean)
• No tool decorator or behavioural contract changed

§8 References — full SQLite-tiebreaker class lineage (instinct)

PR	Surface	Sites	Tiebreaker shape
#30	confidence_log per-row	detect_chains	`id ASC` (integer PK)
#33	confidence_log per-row + indices	history + idx_conflog_*_ts	`id ASC` + DESC→ASC index migration
#34	instincts per-row	recent / list / trending fallback	`rowid DESC` (no integer PK)
#36	instincts per-row + suggest_log LIMIT-1	suggest / search / export_rules / find_duplicates / export_all + _confirm_suggestion	`rowid DESC` (instincts) + `id DESC` (suggest_log)
This PR	aggregate (GROUP BY)	trending real-delta / effectiveness / stats by_category	` ASC`

After this PR merges, all SQLite ORDER BY sites in `src/instinct/store.py` have a deterministic tiebreaker matching their semantic model. The cross-repo rule `feedback_sqlite_rowid_tiebreaker.md` is fully realised in the instinct repo.

🤖 Generated with Claude Code on behalf of Codex (OpenAI)