Merge train: #9827, #9835, #9837

[proggeramlug]

Merge train: #9827, #9835, #9837 — the last three PRs that were blocked on conflicts.

All three conflicted on scripts/gc_runtime_root_holders.json or a ratchet fixture. The inventory conflicts are resolved with a structural merge rather than a text one, because that file cannot be merged textually:

• its entries are JSON objects, so a line-wise union splices one entry's fields into the middle of another (that produced Expecting ',' delimiter twice while I was working these);
• the PASS1_MARKED entry carries source pins, which have exactly one correct value per file — a union will silently restore an older branch's hash over main's, which is how a stale window pin nearly shipped in train Merge train: #9808, #9818, #9819, #9820, #9822, #9823, #9826 #9866.

The merge therefore parses both index stages as JSON, unions entries by (file, name), drops holders that landed PRs have deleted (SHAPE_CACHE_YOUNG, removed by #9756), and always recomputes the window pins from the working tree. With that, #9827 and #9835 resolved first try and #9837 then picked clean.

• perf(gc): route the trigger path through hot TLS — and fix the gate that counts blocks where the cost is per declaration #9827 — route the trigger path and dirty-page barrier through hot TLS.
• perf(gc): pre-size the per-minor dirty-scan covered set instead of rebuilding it from empty #9835 — pre-size the per-minor dirty-scan covered set.
• fix(gc-ratchet): recover main baseline and surface scheduled failures #9837 — recover the GC-ratchet baseline with per-cell evidence, and restore probe liveness so scheduled failures are reported.

Validation

64/64 lint gates; perry-runtime, perry-codegen, perry-hir, perry-stdlib — all green, 0 failures.