A recipe cannot declare its own skip a coverage hole

The mechanism, and why the current behaviour is deliberate

_self_skipped (tools/testmgr.py) reads the recipe's own <target>: SKIP … line and returns it verbatim, so the reason survives into the report in the words the recipe used. skip_summary then asks why.startswith(SKIP_HOLE_PREFIXES) — and why starts with the target name, so it can never match. Not an oversight: the comment above the tuple says folding a recipe's own guard into corpus-absence is the seven-week (corpus absent) bug that tuple was written to end. Do not loosen the match.

Where it is nevertheless wrong

Two different things wear the same shape:

• A recipe guarding its own optional probe. stb_sprintf_probe: SKIP (no library_candidates/stb) — the recipe handles the absence, its jobmates are unaffected, and this is correctly not a hole.
• A recipe declining because the BOX cannot provide coverage. test-zlib: SKIP — gcc oracle not found is the same class as host tool absent: and host dev dependency absent: — the harness's own words for "this box is not providing this coverage". It is invisible in skip_holes, and test-zlib backs one of the two claims this project makes in public copy ("zlib matches the gcc oracle").

What is wanted

A channel the recipe can use to classify its own skip, so the two cases above are distinguishable at the source rather than guessed at by prefix. Shape is open — a second marker line, or a documented reason prefix the recipe emits after the target name, or a hole-declaring variant of the SKIP token. Whatever it is, the harness must keep refusing to infer it.

Provenance and scope

Found while fixing bug-t-corpus-regex-invents-phantom-tree (d6de711d1). NOT the cause of skips: 7 / skip_holes: 2 on the 2026-09-06T18:37:24Z seven report — I attributed the five uncounted gtk jobs to this mechanism and that was wrong. They carry host dev dependency absent:, a HARNESS reason from testmgr.py:1849 that simply is not in the tuple; frankB corrected it and is fixing that separately. None of those seven skips is a recipe self-skip. This ticket is the residual question that survives that correction, with no instance yet observed in a published report — gcc is present on both boxes today.