fp-check: Stop-hook completeness prompt is not path-aware (standard vs deep)

[dguido]

Split out of #190 so the structural hook-scoping fix in #188 can land without losing this.

The gap

#188 moves the Stop hook into skill frontmatter and deletes hooks/hooks.json outright, which fixes hooks firing in unrelated sessions and the main-agent loop. It carries the existing prompt over unchanged, and that prompt demands the full deep-path output for every bug:

Scan the conversation for evidence of ALL of the following for EVERY bug […] Phase 5 (Devil's Advocate): All 13 challenge questions addressed […] Phase 4 (PoC Creation): Pseudocode PoC created […]

A verification that legitimately took the standard path has none of that, so the hook reports {"ok": false} and the run is told it is incomplete when it isn't. A completeness gate that fires on correct output trains people to ignore it.

Proposed fix

@GotenJBZ proposed this in #190 — ordered rules rather than one flat checklist:

1. Skip if the conversation is not fp-check.
2. Skip if the work was delegated to subagents or a workflow (they carry their own per-agent completeness checks, which fix(fp-check): scope completeness hooks to skill/agent frontmatter (#143) #188 puts in each agent's frontmatter).
3. Check only the phases the path actually used — standard vs deep — instead of requiring all five unconditionally.

Rule 2 matters more after #188 than before: with per-agent Stop hooks on data-flow-analyzer, exploitability-verifier, and poc-builder, a delegated run is already checked agent-by-agent, and re-checking at the top level is both redundant and wrong (the coordinator's transcript won't contain the subagents' phase output).

Credit

The diagnosis and the rule ordering are @GotenJBZ's, from #190. #188 was already open with the frontmatter-scoping approach and is CI-validated and version-correct, so it goes in first; this issue is the remainder.

cc @ahpaleus (fp-check owner)