Three workers converged independently on the same split within one hour, which is normally the
signal to act on. This time acting on it would have been wrong, and the reason is measurable.
D, from its scoreboard: what matters is whether the row has been attempted WITH THE CURRENT LEVER
SET, not which pool it is in.
C, from 11 classifications: prio-1 does not distinguish "a name that implies a SOURCE SHAPE" from
"a name that implies a COMPILER BEHAVIOUR".
A, from 14 rows touched: what predicts closability is the WORDING of the row's own note, not its
prio -- and a grep would separate the two groups in one pass.
**The insight is right and is kept. The proposed discriminator is not predictive, and I measured it
rather than arguing about it.** I classified every row A and C named, by outcome, against the pool's
own record:
MATCHED (8): 0x8010036C 0x800690E4 0x80018284 0x80094370 0x80065494 0x800C3514 0x80022E44 0x8002DF1C
-> source-shape 2 | unknown 2 | **residual-class 4**
RESISTED (7): 0x8010400C 0x800A6C34 0x80107CCC 0x80050674 0x80023D40 0x80024630 0x80101E50
-> source-shape 1 | **unknown 3** | residual-class 3
Precision of "residual-class => will resist": **3/7, worse than a coin flip**, and the rule would have
**deprioritised four of the eight actual successes.** The decisive examples are the proposers' own:
C's `0x800C3514` is classed `alloc+layout (priority selector; ...)` and MATCHED -- the lever was
ARITY; B's `0x8002DF1C` is classed `candidate_bytes=104 = CORRECT length, 58 differing bytes` and
MATCHED -- the lever was a `goto` into the body. Both would have gone to the back of the queue.
So `kind` is DISARMED as an ordering rule and retained only as a LABEL written into every pool row,
with the failure and its numbers in the header so no worker trusts it. A test asserts it is not in
the sort key and cannot drop a row, so it cannot be quietly re-enabled.
**This is defect 5's shape exactly** -- a text heuristic over free prose, on this same field, which
inverted 62 rows. The difference is only that this one was tested before being believed. The real
discriminator is whether a residual has been LOCALISED to an allocation or scheduling decision, and
that requires reading the bytes rather than the note. **Free prose is not evidence about closability.**
Also in this commit: `negatives-d-unattempted.tsv`, the 74-row UN-PUSHED band that worker D asked for
and that D's own numbers justify (2-for-7 on rows a previous session pushed to 1-6 bytes, against
one-spelling closes on rows nobody had pushed). Rows with no mechanism named, not registered, not
excluded, not charter-blocked, not held in the ledger. This is D's criterion applied as a band rather
than as a preference.
Pools after the re-cut: 70 rows (a 16 / b 19 / c 21 / d 14).
Christopher Williams
5af450c491