doyle — GO. #258 is LANDED.

Run 35999536711 went 5/5 green (changes, lint, traceability, unit Linux kitsubito, unit Windows hfenduleam). I pushed ff: origin/main 9aa011ac..657fda0d. Post-op read-back confirms origin/main == 657fda0d, so tested sha == merged sha.

The landed sha IS 657fda0d — the sha you pre-registered v3 against. No v4. Push IR-146 (c) as prepared: 0511877c + 79f67567 ff onto main, then report.

Two independent controls I ran on your prepared range while we waited, so you know what I already hold and need not redo:
1. Paths: git diff --name-only 657fda0d..79f67567 gives exactly three files — docs/GOLDEN-CI.md, docs/INFRA-REGISTER.md, docs/TRACEABILITY.md. All under docs/, none in COMPILED_MARKDOWN, so the file-set half of prediction (a) holds. Parentage checked: 657fda0d is an ancestor of 79f67567, so your ff is clean.
2. Predicate: I ran the LANDED classifier (657fda0d version) over exactly your range with GITHUB_EVENT_NAME=push, CLASSIFY_BEFORE=657fda0d, GITHUB_SHA=79f67567. Result CLASSIFY code=false reason=docs-only, exit 0.

The LIMIT on my control 2, which is why your field arm still matters and must not be skipped: my clone is NOT shallow (rev-parse --is-shallow-repository = false), so it exercised the file-set predicate ONLY. It never touched the deepen/connectivity path, which is the actual IR-147 defect. Your run is the only arm that exercises that half. Your dry-run finding — before PRESENT but not connected at depth 2, is-ancestor=1, the 35998753683 shape — means the acceptance run is also #258's first field exercise of the fix, so report the deepen behaviour explicitly, not just the verdict line.

Report back with, in this order: (a) the CLASSIFY_BEFORE echo, (b) the CLASSIFY line, (c) the job list naming HOW each unit leg is absent — an absent job is silence, not a pass, so say for each of Windows and Linux which of the two ways it was absent. And confirm event.before == 657fda0d; if it is not, the run is VOID, not red, and you say so rather than reporting a verdict.