From d7b66a8ed9e18467f0b94f53b3f324390370e799 Mon Sep 17 00:00:00 2001 From: Vasilev Dmitrii Date: Sat, 5 Sep 2026 06:56:51 +0700 Subject: [PATCH] skill: a first-error count is not an unblocking count From #3208. The ungated serde derive was the FIRST rustc error in 84 specs and discriminated perfectly -- 0 of 224 passing files, 84 of 357 failing. Measured by name after the fix, rustc acceptance moved 224 -> 237: +13 unblocked, 0 regressions, and 0 specs still failing on serde. Errors queue. 84 is how many specs the cause stops first; 13 is how many the fix carries to acceptance. Both are true and neither implies the other, so a report needs the first-error count, the by-name before/after, and a count of specs still failing on the cause -- three separate claims. Refs #3208 Co-Authored-By: Claude Opus 5 --- .claude/skills/ci-gates/SKILL.md | 24 +++++++++++++++++++ ...-error-count-is-not-an-unblocking-count.md | 19 +++++++++++++++ 2 files changed, 43 insertions(+) create mode 100644 docs/now/2026-09-05-a-first-error-count-is-not-an-unblocking-count.md diff --git a/.claude/skills/ci-gates/SKILL.md b/.claude/skills/ci-gates/SKILL.md index 6e4304ba90..1a9c93ba55 100644 --- a/.claude/skills/ci-gates/SKILL.md +++ b/.claude/skills/ci-gates/SKILL.md @@ -16005,3 +16005,27 @@ Three rebuilds of the same file in one pass, two of them wrong: one dropped a ma resurrected a withdrawn one. **Both were caught by re-deriving the answer rather than by trusting the previous step** -- `master lost N` and `withdrawn resurrected N` printed after every attempt. A repair that does not verify itself is another edit. + +## 560. A first-error count is not an unblocking count + +A fan-out audit found a real class in the Rust emitter: the serde derive was gated +behind a cargo feature in `gen_struct` and written bare in `gen_enum`, 43 lines below +in the same file. The discriminator was perfect -- the ungated form occurs in 0 of the +224 files that pass and 84 of the 357 that fail -- and two independent verifiers, +one on cause and one on population, failed to refute it. + +The tempting sentence was "this fix unblocks 84 specs." Measured after the change, +by name: rustc acceptance went 224 -> 237. **+13, zero regressions**, and zero specs +still failing on serde where 84 did. + +Both numbers are right and they answer different questions. Errors queue: 84 is how +many specs this cause stops FIRST, and 13 is how many the fix carries all the way to +acceptance. The other 71 simply report whatever stood behind it. A class can be the +largest single cause in a column and still unblock a small fraction of it, and that is +not a failure of the fix -- it is what a backlog with depth looks like. + +Report all three, because they are three separate claims: the first-error count proves +the cause is large, the by-name before/after proves what the fix yields, and a count of +specs still failing on that cause proves the cause is gone rather than merely rarer. +Measure the second one after the change and diff by spec name, so a gain and a +regression cannot cancel in a total. diff --git a/docs/now/2026-09-05-a-first-error-count-is-not-an-unblocking-count.md b/docs/now/2026-09-05-a-first-error-count-is-not-an-unblocking-count.md new file mode 100644 index 0000000000..9be4b81a4e --- /dev/null +++ b/docs/now/2026-09-05-a-first-error-count-is-not-an-unblocking-count.md @@ -0,0 +1,19 @@ +# NOW -- A first-error count is not an unblocking count (2026-09-05) + +Recorded from the serde fix in #3208, where the two numbers differed by a factor of +six and quoting the larger one would have been wrong. + +## The two questions a rejection class answers (Refs #3208) + +- 84 specs had the ungated serde derive as their FIRST rustc error, and the form discriminated perfectly: 0 of the 224 passing files, 84 of the 357 failing ones +- after the fix, measured by name over all 650 specs: rustc acceptance 224 -> 237, so **+13 unblocked, 0 regressions** +- specs still failing on `serde`: **0**, down from 84 -- the cause is gone, not merely rarer +- errors queue: clearing the first cause in 84 specs carries 13 all the way to acceptance and leaves 71 reporting whatever stood behind it +- a class can be the largest single cause in a column and still unblock a small fraction of it; that is what a backlog with depth looks like, not a failure of the fix + +## Why all three numbers belong in the report (Refs #3208) + +- the first-error count proves the cause is large +- the by-name before/after proves what the fix yields, and diffing by spec name is what keeps a gain and a regression from cancelling inside a total +- a count of specs still failing on that cause proves it was removed rather than reduced +- they are three separate claims and no one of them implies another