release: v3.2.0 — fix #833 + ship Unreleased (was: v3.1.1 release)

[robotrocketscience]

Closes #833.

Replaces the prior release: v3.1.1 shape on this branch after the review thread surfaced that tagging current main at v3.1.1 would be a SemVer violation: [Unreleased] includes #814's Removed entry with four explicit breaking-API changes (apply_feedback(propagate=) kwarg → TypeError, FeedbackResult.{pressured_locks,demoted_locks} → AttributeError, aelf locked --pressured CLI flag, aelf:feedback/aelf:confirm MCP payload keys → KeyError). Patch-version bumps cannot ship breaking changes; the SemVer-correct shape is v3.2.0.

What changed vs the prior PR shape

• pyproject.toml + uv.lock: 3.1.0 → 3.2.0 (was: 3.1.0 → 3.1.1).
• CHANGELOG/v3.md:
  • Renames ## [Unreleased] → ## [3.2.0] - 2026-05-15 and opens a new empty ## [Unreleased] above (per top-level CHANGELOG.md policy: "On release, move [Unreleased] content into a dated ## [X.Y.Z] section").
  • Prepends the #833 Fixed entry to the new [3.2.0] ### Fixed section.
  • Updates compare-link footers: [Unreleased]: …v3.2.0…HEAD, new [3.2.0]: …v3.1.0…v3.2.0. (Fixed a pre-existing drift in the same edit — the previous [Unreleased]: …v3.0.1…HEAD should have been …v3.1.0…HEAD after the v3.1.0 ship.)
• Regression test commit (6c8963ef test: regression for #833 demotion_pressure post-DROP read) — unchanged from 3d1aecbb.

What ships in v3.2.0

Section	Issue	Summary
Added	#816	Hot-path touch state v1 storage substrate
Added	#817	ζ posterior-rerank surface (bounded sigmoid contribution)
Added	#758	Seventh meta-belief consumer: adaptive posterior_temperature
Fixed	#833	aelf v3.1.0 crash on any DB previously touched by post-#814 code
Fixed	#834	Auto-install gated to uv-tool installs only
Fixed	#798	Rebuilder pack accounting honours use_type_aware_compression
Fixed	#809	Subfloor noise-pattern filter at sentence-level ingest
Removed	#814	CONTRADICTS-edge auto-demote surface (breaking — see entry for caller migration notes)

Verification

• 2 atomic signed commits (%G? = G); no rewrite of 6c8963ef.
• Rebased onto current main (d3ea7ef5) — clean FF.
• release-docs-check (existing) passes: [3.2.0] section + compare-link footnote exist in CHANGELOG/v3.md.
• release-docs-check (after ci(release-docs-check): assert [Unreleased] drained on release PR (#837) #838 lands) would also pass: [Unreleased] is now empty.
• pytest 3.12 / 3.13 will exercise the tests/test_demotion_pressure_drop.py regression on the rebased head.

Out of scope

• Migration-policy CI guard against future destructive _MIGRATION_STATEMENTS entries without paired reader updates — deferred follow-up.
• Path A (hotfix from v3.1.0 tag, ship v3.1.1 with only [bug] aelf v3.1.0 crashes on any DB touched by post-#814 code — demotion_pressure schema/reader skew #833) — feasible but does not work with the merge-train (FF-only push to main) and produces backport friction since the CHANGELOG was split (feat(ingest): subfloor noise-pattern filter + intra-turn edge-anchor demotion (#809) #810/docs(hibernation): mark memo as design-stage / unimplemented (#825) #832) after v3.1.0. Path B (this PR) is cheaper and SemVer-correct given the breaking changes already on main.

[sourcery-ai]

Sorry @robotrocketscience, you have reached your weekly rate limit of 500000 diff characters.

Please try again later or upgrade to continue using Sourcery

[github-actions]

PR-size soft cap

This PR is over the advisory size threshold:

• 129 changed lines (limit: 200)
• 4 changed files (limit: 3)

Bigger PRs collide with more open work, which under the parallel-session workflow tends to produce repeated attn:merge-conflict cycles (see #602). When practical, split into smaller PRs that each touch a focused surface.

This is advisory only — nothing is blocked. If the size is intentional (large refactor, module removal, generated code), apply the size:override label and this comment will be removed on the next push.

[coderabbitai]

Important

Review skipped

Ignore keyword(s) in the title.

⛔ Ignored keywords (2)

• docs:
• release:

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

Run ID: d92ae1c7-25da-4d47-8898-1e88c162a6ae

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch fix/issue-833-demotion-pressure-skew

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[robotrocketscience]

[claim:review:prince:2026-05-15T01:32:04Z]

[robotrocketscience]

Reviewed e028573d (and prior 3d1aecbb). Reproducing the bug locally is unnecessary — the patch path matches what main already does. Two issues, the second is a blocker.

✅ What's right

• Atomic split. Test commit and release commit are separated — matches the "atomic commits beat batched" rule in ~/.claude/CLAUDE.md.
• Test correctness. _seed_v3_1_0_store mirrors the v3.1.0 column set, then opening with current MemoryStore exercises the DROP COLUMN demotion_pressure migration in _MIGRATION_STATEMENTS. The assertion that get_belief("b833") round-trips without IndexError directly guards _row_to_belief's defensive-keys contract. Second case verifies idempotent re-open. The minor noqa: SLF001 on s._conn.execute(...) for the PRAGMA table_info check is acceptable scope-bracketing for a regression test.
• Version bump. pyproject.toml 3.1.0 → 3.1.1 and uv.lock re-locked.
• CI. All required green: pytest 3.12 + 3.13, CodeQL python + actions, bench-smoke, eval calibration, staging-gate (secrets/pattern/history/release-docs/commit-msg/pr-title/pr-body), deptry, vulture, typos.
• Commits signed. Both 3d1aecbb and e028573d validate (%G? = G).

🚨 Blocker — release scope / SemVer mismatch

The PR cuts v3.1.1 from current main. Current main's CHANGELOG/v3.md [Unreleased] section contains five entries that have landed since the v3.1.0 tag:

• Added: feat(retrieval): H4-only hot-path touch state — boolean rerank multiplier (closes #748) #816 hot-path touch state v1, feat(retrieval): ζ bounded posterior-contribution rerank surface (closes #800) #817 ζ posterior-rerank surface
• Fixed: A4 bench gate is structurally vacuous — rebuilder uses verbatim token cost, ignoring compression flag #798 rebuilder pack accounting honours use_type_aware_compression, feat(ingest): SVO-extractor min-length floor + edge-anchor demotion (#785 §3) #809 subfloor noise-pattern filter
• Removed: refactor(feedback): remove unreachable CONTRADICTS-edge auto-demote path (per #605) #814 CONTRADICTS-edge auto-demote surface — this entry documents four explicit breaking changes (apply_feedback(propagate=) kwarg → TypeError, FeedbackResult.{pressured_locks,demoted_locks} → AttributeError, aelf locked --pressured → unrecognised-argument, aelf:feedback / aelf:confirm MCP payload keys → KeyError)

This PR moves only the new #833 Fixed entry into [3.1.1] and leaves the other five in [Unreleased]. That produces two compounding problems:

1. Project release policy violation. Top-level CHANGELOG.md is explicit: "On release, move [Unreleased] content into a dated ## [X.Y.Z] section and add the compare-link footnote at the bottom of the same file." The release-docs-check job in .github/workflows/staging-gate.yml only checks that the ## [X.Y.Z] header and compare-link footnote exist — it does not check that [Unreleased] was actually drained. That's the gap that let this CI-green.

2. SemVer violation. Tagging v3.1.1 at this HEAD ships every commit in v3.1.0..HEAD, which includes refactor(feedback): remove unreachable CONTRADICTS-edge auto-demote path (per #605) #814 with its four breaking-API changes. A patch-level bump for breaking changes is a SemVer break.

Resolution — two paths

• A — Patch hotfix from the v3.1.0 tag. Branch from v3.1.0 (commit 0b298728), cherry-pick or port the _row_to_belief defensive read + regression test, leave main alone, tag v3.1.1 from the release branch. Smaller blast radius; ships only what users with installed v3.1.0 need. Leaves main's [Unreleased] entries to ride v3.2.0.
• B — Promote main to v3.2.0. Migrate all five [Unreleased] entries plus the new [bug] aelf v3.1.0 crashes on any DB touched by post-#814 code — demotion_pressure schema/reader skew #833 Fixed entry into a dated ## [3.2.0] - 2026-05-15 section in CHANGELOG/v3.md. Bump pyproject.toml to 3.2.0 (minor, not patch — required by SemVer because of the refactor(feedback): remove unreachable CONTRADICTS-edge auto-demote path (per #605) #814 breaking changes). Update PR title accordingly.

A is the cheaper unblock for installed v3.1.0 users (which is what #833 is actually about). B is the right call only if we want everything currently on main to ship now.

Recommend A.

Follow-up — release-docs-check should catch this

release-docs-check should also assert that ## [Unreleased] is empty after the new dated section, or fail. That would have blocked this PR pre-review. Worth a follow-up issue (small docs-job patch).

🔸 Minor observation (not gating)

The PR title release: v3.1.1 — ... and branch name fix/issue-833-demotion-pressure-skew disagree on the conventional prefix. release: is the right title prefix for a version-bumping PR (per ~/.claude/CLAUDE.md). The branch name being fix/... rather than release/v3.1.1 is a minor convention drift but not a blocker — the commit subjects and PR title use the correct prefixes.

---

Not adding ready-to-merge. Leaving attn:review so the next session sees this and routes to either A or B. Release claim on this PR.

[robotrocketscience]

[release:review:prince:2026-05-15T01:35:43Z]

[robotrocketscience]

[claim:review:prince:2026-05-15T02:59:13Z]

[robotrocketscience]

[claim:review:oppenheimer:2026-05-15T03:00:34Z]

[robotrocketscience]

[release:review:oppenheimer:2026-05-15T03:00:38Z]

[robotrocketscience]

Pushed Path B per the review thread. Rationale recap:

1. Tagging current main at v3.1.1 ships #814's Removed entry, which is explicitly breaking (four call-site TypeError/AttributeError/KeyError/unrecognised-argument signatures listed in the entry body). SemVer requires a minor bump → v3.2.0.
2. Path A (hotfix branch from the v3.1.0 tag) is SemVer-correct but doesn't compose with the existing merge-train (which only does FF-push to main), and the post-v3.1.0 CHANGELOG/v3.md split (feat(ingest): subfloor noise-pattern filter + intra-turn edge-anchor demotion (#809) #810 / docs(hibernation): mark memo as design-stage / unimplemented (#825) #832) means a hotfix-branch CHANGELOG edit would need a separate backport-to-main reconciliation pass.
3. Path B (this PR, now) ships the same #833 fix plus the other six entries that already landed on main, all under one SemVer-correct tag, through the normal release infra.

Branch force-pushed onto current main (d3ea7ef5). The regression-test commit (6c8963ef) is unchanged from the prior 3d1aecbb — only the release commit was rewritten. Two signed commits, FF on main.

Re-adding ready-to-merge. Reviewer notes:

• release-docs-check will exit 0 (head 3.2.0 ≠ base 3.1.0 → enters release-PR flow; [3.2.0] section + compare-link present).
• Once ci(release-docs-check): assert [Unreleased] drained on release PR (#837) #838 lands the [Unreleased]-drain assertion, it would also exit 0 on this PR — [Unreleased] is now empty.
• If we'd merged the prior v3.1.1 shape, both the project policy and SemVer would have been violated; the new [Unreleased] assertion in ci(release-docs-check): assert [Unreleased] drained on release PR (#837) #838 specifically catches that failure mode.

[robotrocketscience]

[release:review:prince:2026-05-15T03:06:18Z]

[github-actions]

merge-train: blocked

branch is not fast-forward on main (branch base d3ea7ef5e9af80ca02cbd391a42e1169f05f6b38, current main 75d27b38a6a0838a8b8868aa9afe10de05fa57ac). Rebase locally (git rebase github/main), force-push, and re-add the label.

The ready-to-merge label has been removed. Address the issue above and re-add the label when you're ready for another attempt.

[github-actions]

merge-train: merged 840a792 → main via FF push.