Docs: record release-plan status at 2026-07-29

[PurHur]

The plan still reads as though nothing has landed. This records where things actually stand.

Status

item	state
P1.1 CI on lib/, ext/	done — compiler-gate.yml (#24275), green on master
P1.3 suites runnable	done — shard-compliance.sh (#24445) + compliance-baseline.sh (#24469)
P1.4 flaky quarantine	mechanism done, list deliberately empty until two-run evidence
P2.5 side-loadable extensions	#24285, #24387, #24418, #24421
P3.9 spine split-TU	blocked, root cause mapped (#24429)

Three corrections measurement forced

1. Phase 2.5 buys nothing for user binaries. Unused extensions are already never linked — a
   hello-world built with 76 extensions vs 6 differs by 360 bytes. The justification is
   extensibility and the compiler's own build cost, not user binary size.
2. "~20 min sharded" assumes parallel runners. On one box shards run sequentially, so ~2 h.
3. "Clone and compile hello world" was ~9 minutes, not seconds — the corpus warmup ignored the
   committed cache. Now 5 s, and gated so it cannot drift back.

The pattern worth carrying forward

Nearly every problem found this week was a check that reported success without executing:

• a hardcoded emit_path=native success string;
• a precondition gate green through an entire outage;
• external_method_stubs=0 from a build that aborted before the report ran;
• grep -c … || echo 0 silently disabling its own zero-executed guard;
• a timed-out shard writing a .failed file that looked complete;
• a partition reshuffled mid-run by a git pull.

None was a wrong answer. All were absent answers wearing a clean result. Three were in code I
wrote this week, which is why the habit that matters is: validate a gate by making it fail, and
distrust any zero or clean result until you know the check ran.

Not merging yet

A clean 24-shard baseline run is in flight. Merging needs a pull, which is the exact hazard #24537
documents. This sits until the run finishes.