Diagnostic Legibility: invocation-persistence corpus for Phase-C escalation trigger

[russmiles]

Background

The Phase-C-escalation trigger introduced in sub-S2b needs a corpus of diagnostic-legibility agent invocations to compute its signal (the sentinel-only-ratio that would justify promoting Phase C cross-check to a separate agent). That corpus has no owner:

• S2b deferred building it.
• S3 (§8) named S4 as its natural home.
• S4 (the /diagnose command, Diagnostic Legibility — S4: On-demand surfacing interface for human legibility #333) deliberately scoped it out: /diagnose writes one report per run and accumulates nothing.

With parent #327 closing on S4 ship, this concern loses its natural tracking home — hence this standalone issue.

What this is

A decision-archaeology finding promoted from the S4 choice-cartographer record (Story #8, "Observability scoped out where S3 pointed it home"): a "natural home" hand-off in slice N does not bind slice N+1, and a concern handed forward three times can fall through the gap. This issue is the re-filed tracking home.

Scope to decide (not yet designed)

• Whether /diagnose (or the agent dispatch path) should append a per-invocation record to a corpus.
• The corpus shape, location (cf. the derived-artefact .gitignore stance — S4 cartographer Story Sync: add framework-change-prompt Stop hook #2), and retention.
• The escalation trigger's exact signal and threshold.

This is a future-optimisation signal, not a correctness mechanism — no failure is currently undetected. Prioritise only when re-run cost or escalation evidence makes it worth the design.

References

• S4 cartographer record: docs/superpowers/stories/dl-s4-diagnose-command-design.md Story Document harness observability in README #8
• S3 §8 (the hand-off); S4 spec §8 / §10.6 (the decline)