twatch: suppress downstream auto-filing when a root job (fpc-bootstrap) is red

• Type: feature — Track T (watcher behavior). Opened by fable-O after the 2026-07-18 cascade.
• What happened: one missing FPC-seed forward turned fpc-bootstrap red; the next watcher pass auto-filed 939 regression tickets — every FPC-dependent job (lib-fpc-clean, test-asm suite, FPC-built test-core sweep) filed its own. One root cause, 939 files of noise (bulk-swept to rejected/, see [[regression-cascade-f5c8fbec-fpc-bootstrap]]).
• Ask: teach twatch job dependencies (or at least the one big edge: fpc-bootstrap red ⇒ fold all FPC-dependent reds into the fpc-bootstrap ticket as a cascade list instead of filing each). A red ROOT job should file ONE ticket naming the cascade set.
• Gate: T tooling gate (tools/testmgr.py --tier full green) + a scratch bare-repo dry run of the filing path.

Resolution (fable-O, same evening — user-directed)

Implemented as a threshold, not a dependency graph: file_stub_tickets now routes any sweep with > CASCADE_THRESHOLD (10) new reds to file_cascade_ticket — ONE ticket keyed on the bad sha (dedup via the same any-bucket check), listing the full job set, naming root-cause suspects when fpc-bootstrap/selfhost-fixedpoint are in the set. Small sweeps keep per-job stubs. Scratch-harness verified: 3-red sweep → 3 stubs; 50-red sweep (incl. fpc-bootstrap) → 1 cascade ticket with the root named; same-sha rerun → no duplicate. Full dependency modeling stays future work if the threshold ever misroutes.

Log

• 2026-07-18 — resolved, commit 70b16286.