fix(queen): the Queen's own slot numbers made her unable to see the answers

[gHashTag]

The finding

153 dispatches sit in sendBack having spent no retry budget. Not one has a single criterion that was tested and failed. Every one carries a complete, parseable ## VERDICT block. They were returned for criteria they had already answered.

The cause

The brief hands the bee numbered slots and asks for - 1. <criterion>: met, so parseVerdictBlock returns a criterion beginning "1. ". The promised text carries no number.

unjudgedCriteria documents matching by containment in either direction:

a criterion counts as judged when one side contains the other — containment in EITHER direction, because the 300-character slice means the promised text may contain the line's text and not the other way round

One of those two directions could not fire even once. want.includes(line) is impossible when the line always carries a prefix the promise lacks. Only line.includes(want) survived, and that demands the bee reproduce the criterion whole. Shorten it by one trailing sentence and the criterion reads as never answered.

Because "unanswered" is deliberately not charged against the retry ceiling (browseros-ai#1420 FR-003), these never escalate and never stop. That is the 80-issue ceiling-less loop reported in #471, and this is its cause.

The measurement

Re-judged all 153 with the review's own functions, against the live board:

no verdict lines at all : 0
every criterion judged  : 0
some unjudged           : 153
   76  promised=5 verdictLines=5 unjudged=1 failed=0
   31  promised=4 verdictLines=4 unjudged=1 failed=0
   27  promised=5 verdictLines=5 unjudged=2 failed=0
   ...
unjudged>0 with the CURRENT matcher        : 153
unjudged>0 after stripping the "N. " prefix:   5
fixed by stripping alone                   : 148

148 of 153. The remaining five are unjudged for reasons of their own.

Real example (browseros-ai#1527) — promise vs. what the bee wrote:

promised: `LC_ALL=C grep -cP ...` prints 0, and the raw output is quoted in the
          report. The command MUST NOT name or enumerate the specific items ...
wrote   : 1. `LC_ALL=C grep -cP ...` prints 0, and the raw output is quoted in
          the report

The repository already had the right answer

missingVerdictSlots strips the number and matches by slot. Run over these same 153 it reports every one complete. Its own comment claims it and the review "read the SAME block ... so a line this counts as answered is exactly a line the review counts as judged." That was false — two implementations of one rule, disagreeing, with a comment asserting they could not.

Why the suite missed it

There was already a case for each direction of the match, including the truncated-quote case. Every one used an unnumbered criterion, and production has not sent an unnumbered one since the brief began numbering slots. A numbered line that quotes its criterion in full still matches on the old code, so only the shortening shape fails: the numbered cases were never written, and the shortening cases were never numbered.

The change

One line, inside normalize, applied to both sides so this stays one rule rather than a special case at a call site. The filter returns the original strings, so nothing a person reads loses its number.

Verification

Extracted the real unjudgedCriteria from the committed blob and ran it directly (the function is self-contained, so it needs none of the server's deps — this checkout cannot run the full suite):

• before the fix: 2 fail, 5 pass
• after the fix: 7 pass, 0 fail

The two failing cases are the discriminating ones. The other five are labelled guards and pass on both sides, so the strip cannot be bought at the cost of the containment fence — in particular a short numbered line still does not judge a long criterion, and a criterion opening with a decimal keeps its meaning.

CI is the witness for the full suite.

🤖 Generated with Claude Code

[github-actions]

✅ Tests passed — 2064/2095

Suite	Passed	Failed	Skipped
✅ agent	87/87	0	0
✅ build	9/9	0	0
✅ cdp-protocol	5/5	0	0
✅ eval	93/93	0	0
✅ server-agent	272/272	0	0
✅ server-api	951/981	0	30
✅ server-browser	6/6	0	0
✅ server-integration	10/11	0	1
✅ server-lib	273/273	0	0
✅ server-pglive	3/3	0	0
✅ server-root	68/68	0	0
✅ server-skills	31/31	0	0
✅ server-tools	242/242	0	0
✅ shared	14/14	0	0

View workflow run

[gHashTag]

Correcting my own claim in this PR

The description says "The repository already had the right answer and did not know it", meaning missingVerdictSlots. That is wrong, and I'd rather correct it than have someone act on it.

Measured across the whole board (358 verdict blocks with criteria):

blocks whose lines carry a N. prefix	202
blocks with no numbering at all	156
…of those, where slot matching finds nothing covered	156 of 156

So slot matching is blind on 44% of the board. Wiring it as the review's matcher would mark every criterion in those 156 unanswered — this PR's defect with a wider blast radius.

The two functions disagree on 347 of 358 rows, in both directions:

  76  sendBack: bySlot=5 byText=4 of 5    <- the defect this PR fixes
  44  accept:   bySlot=0 byText=4 of 4    <- slot matching blind on unnumbered
  34  accept:   bySlot=0 byText=5 of 5
  32  sendBack: bySlot=4 byText=3 of 4

Neither is right alone. Text matching was blind to the number; slot matching is blind to its absence. Stripping the number and then matching by text is the only rule that reads both the numbered briefs and the unnumbered ones — so the fix is a strip, not a swap.

The diff is unchanged; this adds two comments:

• a DO NOT SIMPLIFY INTO SLOT MATCHING note at the call site, because the resemblance is exactly what makes the wrong move attractive later
• removal of the false sentence in missingVerdictSlots: "one parser, so a line this counts as answered is exactly a line the review counts as judged". Sharing a parser is not sharing a rule, and that sentence is what made this defect look like somebody else's already-solved problem.

Found by building the agreement gate that this round set out to write — the gate's first run corrected the round that motivated it.