Retain the 03_ingest emission board at the first main SHA whose floor reports failed=0

[gunbai-bot]

Pins and retains the emission board at 907f19c2cc7 — the first main SHA whose required floor reports failed=0 (run 32633501354, planned=10672 executed=10672 terminal=10672).

Why now, and why the log rather than the counts

The board has been unmeasurable for most of today. A §4c annotation inside a match body refused the compiler at strict preparation from ~04:30 until #8989 landed, so anything measured in that window carries no information about emission — and, being a refusal upstream of the fold, it produced clean-looking output rather than obviously broken output. Main is now repaired: #8998, #8987 and #8992 partitioned the eight width failures disjointly, and nothing had verified they compose until that run.

The log is retained, not just the histogram. A board gets repartitioned many times; without the artifact each pass rebuilds the compiler and re-emits, paying the expensive transaction once per iteration instead of once. With it in the tree, later classification is offline analysis over a fixed artifact — and the artifact carries the SHA that produced it, so a count cannot drift from its subject.

The board

177 files emitted, 503 diagnostics, cargo refuse
coded rustc errors (direct grep of the retained log): 316
E0308:128  E0425:24  E0599:23  E0277:23  E0004:21  E0609:18
E0560:17   E0061:17  E0631:9   E0433:8   E0369:7   E0614:6
E0282:6    E0071:3   E0728:2   E0310:2   E0533:1   E0223:1

The identical result, and why it is not a false green

This matches the certified board at 98b18cdc — same 316, same histogram position for position. That is the shape this program has learned to distrust first, so the discriminators are recorded rather than the conclusion asserted:

• PROV_BIN_BEFORE=0 — the compiler was rebuilt from this tree, not reused. A cached binary reporting a new SHA is precisely how a stale board reads as a fresh one.
• The probe row carries this SHA.
• The log bytes differ across all three runs taken today — 234165 / 234279 / 234163, three distinct sha256 — while the error population does not. A run that did not execute cannot produce a byte-different log with the same population.

The merges between those SHAs were witness, fleet, roster and prose work; none touched the emitter. So an unchanged 03_ingest emission is the expected result, and it is now measured rather than assumed.

Units

The README states which of four available counts answers which question: 503 is the emitter's diagnostics including warnings and notes; 316 is coded rustc errors and is the board number; 330/331 are the probe's histogram sums folding in the uncoded rows. Quoting one where another was measured is the unit error that made two earlier boards in this program unsound.

What it is for

Classification transfers onto this log. A prior classification carries forward only where the load-bearing identity still joins — emitted module, enclosing declaration, callee or field identity, error code or a declared successor code, normalized expected/found. Generated line and column are evidence, not identity. A historical row absent from this log gets no further analysis, which is what makes board-first cheaper than finishing the old partition rather than more expensive.

Docs only: one log and one README. No code, no carrier changes.

— sent from smart-ram-730