feat(loop): free a retry ceiling that was spent on silence

[gHashTag]

The swarm ran one bee of four. This is why, and what it took to get to three.

Nine issues parked, and not one of them on the work

The tick's own skip summary said claimed: 11, and nine of those were dispatches at or over QueenRetryPolicy.maximumRealAttempts = 2 — the oldest idle 52 hours. Meanwhile author.mjs found fourteen real deficits and could file against none of them, because every one already had an issue: one of the parked ones. Work existed and was locked by the thing that recorded its failure.

Then the review notes, read together:

issue	verdict	omission
browseros-ai#1133	4 criteria still unmet	omitted 4 verdict lines
browseros-ai#1175	4 criteria still unmet	omitted 4 verdict lines
browseros-ai#1316	3 criteria still unmet	omitted 3 verdict lines
browseros-ai#1318	4 criteria still unmet	omitted 4 verdict lines
browseros-ai#1311	5 criteria still unmet	omitted 5 verdict lines

The unmet count is the omitted count. Those criteria were not judged and found wanting — they were never judged at all. The reviewer marks a criterion unmet when the worker's ## VERDICT block has no line for it. That is the right default (silence must never read as success) and it is a different fact from a criterion that was tested and failed.

So the whole retry budget was spent on the shape of a report.

What unpark.mjs claims, and what it refuses to

maximumRealAttempts counts real attempts. An attempt that returned no verdict on a criterion did not attempt that criterion.

It releases only those, and the test is deliberately strict — every unmet criterion must be accounted for by an omitted line:

FREE  #1175   escalate  sb=2   40.9 h  all 4 unmet criteria accounted for by 4 missing verdict line(s)
FREE  #1311   sendBack  sb=2   40.4 h  all 5 unmet criteria accounted for by 5 missing verdict line(s)
FREE  #1133   escalate  sb=2   39.6 h  all 4 unmet criteria accounted for by 4 missing verdict line(s)
keep  #1291   escalate  sb=2   39.3 h  6 unmet against 5 omitted - at least 1 criterion was tested and failed
keep  #1328   sendBack  sb=2    4.1 h  the review names no omitted verdict lines
keep  #1329   sendBack  sb=2      1 h  the review names no omitted verdict lines
keep  #1350   sendBack  sb=2    4.1 h  the review names no omitted verdict lines

Three released, four kept. It also refuses anything whose issue body asks for a person, on the same grounds as stale-escalations.mjs.

A filter of mine that could delete the answer it was filtering

clean() stripped railway's chatter by dropping any line containing Migrate, Existing, Using SSH or railway.json. A single-line JSON answer carrying a review note that mentioned migration was deleted in full, and the caller reported "unparseable answer" about a query that had worked perfectly.

Patterns are now anchored to the start of the line, and a line beginning with [ or { is never dropped. A noise filter that can eat evidence is worse than no filter: it turns a working system into an unexplainable one, and it fails silently by construction.

Result

0 → 3 bees. The fourth slot is held by a genuine fileConflict — two of the four issues I filed name the same file, which is an authoring mistake and not a defect. To fill N slots you need N issues with disjoint boundaries.

The skill gains what the night measured

• an escalation is a claim about a cause, and a cause can be measured again — unlike a wait, whose input can never change
• two gates before retiring one, and the near miss that built the second
• ask the shipping parser, never a copy
• what a judging pass actually found: 187 criteria, 0 fabricated by any worker, and 6 of my own 7 accusations refuted on adversarial review