release: add behavioral feature-gate coverage

[samo-agent]

Summary

Fills the per-feature behavioral release-gate gap documented in #160 and absorbs the complete release-gate foundation developed in #188.

• adds deterministic fixtures and exact assertions for every installed non-underscore public surface
• covers readers, lifecycle/admin side effects, reader grants/revokes/default pg_monitor, documented reader edges, and pgss text enrichment/degradation
• runs the canonical feature suite with both optional extensions, without pg_cron, without pg_stat_statements, and without either
• includes release: add Docker release gate and normalize config upgrade #188's pg_cron Docker images, runner, shared SQL assertions, attnum-ordered schema snapshot, and ash.config ordinal normalization
• retains release: return processed grains from rollups #194's processed-grain contract: empty minute grains count as processed, caught-up calls return zero, and physical rows remain independently asserted

Refs #160, #188, #191, and #194.

Final integration

Hand-rebased onto current main 2ccdf65 after #194. The two stacked #194 commits were dropped: the test patch was recognized as upstream, and the implementation duplicate differed only at the migration header where main carries the corrected, non-contradictory wording. Range-diff preserves every other PR commit.

Three full-workflow findings were fixed without changing product behavior or weakening an assertion:

1. DROP/CREATE EXTENSION pg_stat_statements does not clear its preloaded shared state. After thousands of earlier workflow statements, the exact marker could be evicted. The feature driver now calls pg_stat_statements_reset() immediately after recreating the extension.
2. H-CI-3 launched three synthetic pg_sleep(120) clients but retained neither backend identities nor client PIDs. The later lifecycle fixture correctly observed those three plus its own sleeper. H-CI-3 now tags only its workload, terminates those tagged backends, and reaps all three clients on success and failure. The exact active_count = 1 and decoded-row assertions are unchanged.
3. PostgreSQL 19 beta 2 resolves text || jsonb in two ASSERT diagnostic expressions as JSON concatenation. Both coverage objects are now explicitly cast to text. The asserted bounds and minutes_with_data predicates are unchanged.

#188 absorption and assertion conservation

All original 42 #188-owned shared assertion sites remain and are reachable through the canonical runner. Their files now contain 44 sites after adding the current-main failure-counter and trigger-preservation checks.

Using the established ^\s*assert\b source-site convention across the workflow and devel/tests/*.sql:

• current main before: 858 inline + 0 shared = 858
• this PR after: 848 inline + 145 shared = 993
• 11 existing inline assertions moved into shared SQL; the other 847 remain inline
• 1 new inline plus 134 other net-new shared sites yields +135 overall

No assertion was lost. At #188's original extraction point, its own before/after inventory was 594/594 workflow assertions.

RED/GREEN evidence

• Ordinals: current main's dev-upgrade leaves ash.config.num_partitions at attnum=11; the new assertion exits 3. This head normalizes canonical attnums 1–22 (num_partitions=3) and passes preservation plus idempotent reapply checks.
• rollup_minute() and rollup_hour() return per-database row counts, not processed grains #191: on exact head 08b5f6b, the four feature-bearing surfaces (features, degraded-no-cron, degraded-no-pgss, degraded-neither) all PASS on PG14 and PG17: 8/8, with no Expected FAIL: #191 remaining.
• pgss isolation: after prefilling PG17 pgss to 4,910 entries/17 deallocations, the old no-reset step executes the marker then fails at features_readers.sql:1400; the reset passes the unchanged enrichment assertions.
• H-CI-3 isolation: old head 041a450 fails every cron-on matrix job with four live sleepers. An isolated PG17 check of the fix removes all three tagged backends and tracked clients while leaving a differently tagged control active.
• PG19 diagnostics: old head 61a7547 fails on PG19beta2 with invalid input syntax for type json at the unchanged B5 assertion; exact explicit casts make the canonical PG19beta2 fresh-install surface PASS.

Validation

On exact head 08b5f6b:

• PG14/PG17 four-surface feature gate: 8/8 PASS; no Expected FAIL: #191
• PG19beta2 canonical fresh-install surface: PASS
• PG17 canonical upgrade chain: PASS
• fresh install, full dev-upgrade, and actual v1.5-to-current snapshots: byte-identical
• config row/default/owner/ACL/constraint/index/self-FK/trigger preservation and OID-stable reapply: PASS
• release: add Docker release gate and normalize config upgrade #188 shared assertions: 42/42 retained; final inventory 993 static sites
• bash -n devel/scripts/release_gate.sh: PASS
• release-gate helper unit tests: 7/7 PASS
• workflow parse, extracted Bash syntax, Python AST parse, and git diff --check: PASS
• independent Codex and focused cleanup reviews: no merge-blocking finding

Final landing rebase

The merge candidate is hand-rebased onto current main at 9efad9c (after #198 and #199). Exact head: fcd37d7.

• GitHub Actions has a 512,000-byte workflow-file limit. Combining release: add behavioral feature-gate coverage #192 with the new release: keep stale rollups honest without losing fast path #199 inline selector gate would have produced 513,005 bytes, so the unchanged 110-line/9-assert selector transaction was moved byte-for-byte to devel/tests/stale_rollup_source.sql; the workflow is now 508,364 bytes and invokes it in the same psql session.
• The final assertion inventory is 848 inline + 187 shared = 1,035. All 42/42 original release: add Docker release gate and normalize config upgrade #188-owned shared assertions remain; those files now contain 44 sites after later additions. The release: keep compacted query attribution honest #198 query-attribution gate remains 30 assertions and the release: keep stale rollups honest without losing fast path #199 source-selector gate remains 9.
• Exact PG14/PG17 hard gates passed features, degraded-no-cron, degraded-no-pgss, and degraded-neither: 8/8 PASS, with no Expected FAIL: #191 in source or output.
• The first exact combined run correctly failed because release: keep compacted query attribution honest #198 added the dynamic reader helper ash._exact_query_uses_coarser(integer,integer,name), taking the direct reader-function count from 43 to 44. The final gate expects 44 and also names that helper explicitly for both the feature reader and pg_monitor, so an unrelated 44th grant cannot mask its loss.
• A deterministic fractional-boundary fixture exposed a real ash.report bug: convert-then-floor rounded :59.9 endpoints into the following minute. The implementation now floors each bound before epoch conversion. Parent RED shifted both bounds by one minute and reported 4 expected/3 with data; final GREEN covers the exact four-minute range. This user-visible fix is recorded in RELEASE_NOTES.md; no assertion was weakened.
• Fresh-install and upgrade-chain release gates pass on PG17; workflow YAML, shell syntax, helper tests (7/7), and git diff --check pass.

This section supersedes the historical base/head and assertion totals above; those remain as the audit trail for the earlier review point.

[NikolayS]

Codex review standing in for REV — audit verdict: PASS on exact head 041a450. Hand rebase onto main 2ccdf65 correctly dropped the two stacked #194 commits; the only stopped conflict was their obsolete migration-header wording, for which current main is the corrected superset. Range-diff preserves the other 12 commits. Ordinal RED/GREEN reproduced (main attnum 11 and assertion exit 3; head canonical 1–22 with preservation/reapply PASS). Both schema-equivalence paths are byte-identical. The mandated #191 feature gate is 8/8 PASS across four surfaces on PG14 and PG17. All 42 #188-owned shared assertions remain; source-site inventory is 858 before and 993 after (848 inline + 145 shared), with none lost. Old CI's pgss-marker failure was independently reproduced under saturated shared state and fixed narrowly by reset-after-create; enrichment assertions are byte-identical and pass. Codex static review found no blocker. APPROVED.

[NikolayS]

Final exact-head verification on fcd37d71efbe5c2b86b2374b52a07d42194db4b8: APPROVE.

• Hand-rebased onto exact current base 9efad9c. Range-diff maps all 15 original #192 commits patch-identically; only the documented selector extraction, report-boundary fix, and helper-grant assertions follow.
• Exact-head CI is green across PG14–PG19beta2, docs, and cron-off: https://github.com/NikolayS/pg_ash/actions/runs/30315785772
• Required hard gate is 8/8 PASS: PG14+PG17 × features/no-cron/no-pgss/neither, with zero Expected FAIL: #191.
• Parent RED proves convert-before-floor shifted both :59.9 report bounds +1 minute and returned 3/4 covered minutes. Final GREEN floors before conversion, preserves all four minutes, and release-notes the behavior change; no assertion was weakened.
• The #198 dynamic reader helper made the correct direct function-grant count 44. The final gate asserts both that exact count and named EXECUTE on _exact_query_uses_coarser(integer,integer,name) for the explicit reader and pg_monitor, while admin negative controls remain.
• Assertion inventory is 848 inline + 187 shared = 1,035. All 42/42 original #188-owned shared sites execute on passing paths; those files now contain 44 sites.
• #198 is preserved (30 query-attribution assertions); #199 is preserved byte-for-byte after extraction (9 source-discriminating assertions and the exact least(until, date_trunc(minute, now())) clamp).
• Fresh/dev-upgrade/actual-v1.5 schemas are byte-identical; ordinal, reapply, and OID-stability checks pass.
• Workflow size is 508,364 bytes, below GitHub Actions’ 512,000-byte limit. Shell/YAML/Python syntax, helper unit tests, conflict-marker scan, and diff check pass.

No merge blocker found.