Decide: re-pin now after the dyn-array aliasing flip, or wait for the matrix?

• Type: decision — Track U
• Opened: 2026-08-06 by the A+N session that landed 937c51dc2.
• Why escalated rather than decided: make pin is the one action CLAUDE.md calls "the deliberate brake", and this is a semantics flip. Choosing either way imposes a cost on other lanes, so it is not an agent's call.

The situation

937c51dc2 removed the dynamic-array copy-on-write so b := a aliases, matching FPC. It changes emitted code, and tools/gate.sh quick checks the self-host fixedpoint by seeding stage A from stable_linux_amd64/default/pinned:

sh-A = pinned(compiler.pas)   <- OLD codegen, still emits the COW calls
sh-B = sh-A(compiler.pas)     <- new codegen
sh-C = sh-B(compiler.pas)
pass iff A == B == C

sh-A therefore comes out 138KB larger than sh-B == sh-C, and the step reports FAIL. The real fixedpoint is intact — seeded from the new compiler, A==B==C is byte-identical (6398664 bytes), and make compiler/pascal26 converges — but every lane's gate.sh quick will read RED until pinned is refreshed.

This is inherent to any codegen change, so it is presumably a recurring situation, not new; the chore(stable): pin vNNN commits suggest re-pinning is the routine answer.

The fork

1. Re-pin now (make stabilize → make pin, commit stable_linux_amd64/**). Restores a meaningful gate for everyone immediately. Cost: Track B's ground becomes a compiler with the new dyn-array semantics before Track T's full matrix has swept it. If the matrix finds RTL fallout, B is building on it.
2. Wait for the matrix, then pin. B's ground stays proven. Cost: a red fixedpoint step in every lane's gate for as long as it takes, which trains agents to ignore a red gate — the worst possible habit for the one check that is safe-by-construction.
3. make stabilize only (moves latest, NOT B's ground). Records the checkpoint but does not fix the gate reading, since gate.sh seeds from pinned. Halfway, and does not buy the thing that hurts.

Recommendation

(1), re-pin now, with (2) as the answer if the user would rather hold B still. The evidence behind that: the change is verified against the FPC oracle on x86-64 / i386 / arm32 / aarch64 (flat, nested, managed elements, record field, SetLength-detaches, Copy-detaches), the entire dynarray/open-array test corpus is green (39 tests, zero failures), and testmgr --tier quick is green. The unswept surface is lib/rtl + compiler/builtin — the 63 dyn-array variables and 79 candidate whole-array assignments the original ticket scanned as an upper bound and never resolved per-scope. That is real, and it is also exactly what pinning would surface fastest, via Track B's own gates.

A cheaper hedge if (1) feels early: pin, and treat the first lib-test/demos red as a revert-the-pin trigger rather than a fix-forward.

Not part of this decision

The code is already pushed and T can see it. Nothing is blocked on this except the gate's readability, so a slow answer costs only that.

2026-08-07 — RESOLVED: option 1, re-pinned. v246.

User's call: "go ahead and re-pin."

Done, and the evidence was better than this ticket could assume when it was written. Track T's full tier ran at exactly HEAD 398e72d2c and came back GREEN, and its recorded compiler_sha256 (85203cdaf90eb14f9df8e55851a5989664cc27741f9a10b3e43633ec7da85cd4) is byte-identical to the binary make stabilize recorded as v246. So the pinned compiler is not "a build that passed something" — it is the same bytes the full matrix swept. The wait-for-the-matrix option (2) had effectively already happened by the time the answer arrived.

Verified before committing, in the order that matters:

gate.sh lib was the real risk and the reason the hedge in the recommendation existed: lib/rtl + compiler/builtin hold the 63 dyn-array variables and 79 candidate whole-array assignments the original ticket scanned as an upper bound and never resolved per-scope. They are green, so that upper bound was what it was labelled as — collisions, not findings.

The revert-the-pin trigger named in the recommendation (first lib-test/demos red) is therefore unused, and stands as the trigger if a later matrix run disagrees.

Log

• 2026-08-07 — decided, commit 758fa3936.