fix(release): CHANGELOG.md renderer splice contract + 44 k-line drift sweep (ADR-0913)

[lusoris]

Summary

scripts/release/concat-changelog-fragments.sh used ^## [^[] as the "end of Unreleased block" sentinel. 84 of 102 in-tree fragments contained ## headers in their bodies (the redundant section name ## Fixed or a per-PR descriptive title like ## Vulkan submit-pool migration); each tripped the sentinel and the next --write preserved the post-trip tail as if it were release history. Each cycle inflated CHANGELOG.md by ~3 kB. Master tip 544299fae1 had drifted to 59 757 lines (~95 % duplicated). PR #332 / #383 / #401 all flagged "pre-existing drift" without rooting it; PR #384 / ADR-0892 identified the section-hygiene angle but explicitly deferred the regeneration.

Root cause: the awk pattern in the splice passes used ^## [^[] as the boundary, which matched any ## Foo except ## [Foo]. Fragment bodies legitimately contained ## Foo shapes; the regex was lying about its invariant.

Fix: boundary regex anchored on ^## \[ — release-please's only header shape (## [vX.Y.Z] - YYYY-MM-DD and ## [Unreleased]).

Per-issue fix count

Issue	Count
Renderer awk-boundary regex	1 (both passes share BOUNDARY_REGEX)
Fragments with stray ## headers normalised	84
Fragments with redundant ### Section first-line stripped	18
Fragment dirs outside Keep-a-Changelog set relocated	2 (perf/, performance/)
Fragments renamed perf-<topic>.md under changed/	32
Renderer edge cases added	3 (empty-skip + unknown-dir-warn + h1/h2 demoter)

CHANGELOG.md diff size

Metric	Pre-fix	Post-fix	Delta
CHANGELOG.md lines	59 757	15 050	−44 707
^## headers in file	155	1	−154
--write idempotency	broken	clean	—

ADR-0108 deliverables

• Research digest — docs/research/0913-changelog-renderer-and-drift-2026-05-31.md
• Decision matrix — docs/adr/0913-changelog-renderer-splice-contract.md §Alternatives considered
• AGENTS.md invariant note — scripts/AGENTS.md: splice contract plus fragment hygiene
• Reproducer / smoke-test command — Research-0913 §Reproducer plus test plan below
• CHANGELOG fragment — changelog.d/fixed/changelog-renderer-and-drift.md
• Rebase note — docs/rebase-notes.md (in-flight-branch impact plus resolution)
• state.md row — Recently closed § T-CHANGELOG-RENDERER-SPLICE-AND-DRIFT-2026-05-31

Test plan

• bash scripts/release/concat-changelog-fragments.sh --check → exit 0 (no drift)
• bash scripts/release/concat-changelog-fragments.sh --write && bash scripts/release/concat-changelog-fragments.sh --check → idempotent
• Inject fake ## [3.0.0-lusoris.1] - 2026-06-01 after Unreleased + re-run --write → release-section preserved unchanged
• pre-commit run --files <touched> → all checks pass
• Required Checks Aggregator (will run after gh pr ready)

Relation to PR #384 / ADR-0892

PR #384 introduced the changelog.d/perf/ → changelog.d/changed/perf-*.md migration and the revert/security/style Conventional-Commits additions to release-please-config.json. It explicitly deferred the CHANGELOG.md regeneration sweep. This PR completes the deferred work + roots the actual renderer bug PR #384 missed. If PR #384 lands first, this branch's perf/ renames become a no-op on rebase. If this PR lands first, PR #384's renames become a no-op.

Out of scope

• A --lint mode that fails on stray ## in fragment bodies (render-time demoter is enough; lint is incremental polish).
• The pre-existing docs/adr/README.md drift (201 rows have no fragment-file backing under _index_fragments/; that's a separate back-fill PR).

🤖 Generated with Claude Code