Field-hardening + project-capability routing + a prose readability pass

[rmasl]

Grown from real runs, in three layers — atomic commits, each with its DECISION_LOG entry.

1. Field-hardening (original scope)

• Worktree init resolves a default in-skill instead of blocking: worktree.baseRef self-heals to head; a missing ## Worktree init section is synthesized from .gitignore + the verify commands; sequential-only is the last resort.
• One-page digest ships alongside the full spec (specs/<slug>-digest.md): goal, load-bearing decisions with one-line whys, build order, open risks — each item referencing the full spec's section. Digest first at the review gate.
• Cold-start flow trace is mandatory in both critique prompts + a spec-critique exit gate: walk each primary flow from an empty state and check every step's prerequisites exist at that point. Catches bootstrap/ordering contradictions section-consistency review structurally misses.
• Phase 3 hardening: breakdown parallelizes independent authoring (with the shared-verify-environment caveat), gates re-run through the project's documented harness (an MCP tool named in CLAUDE.md beats a raw-CLI wrapper), and a hard test-rerun scope — touched/failing tests while fixing, the full suite exactly once at the end.

2. Lead with the project's own capabilities (new)

Discover once, as a precondition, what the project ships — skills, read-only research/"hunter" subagents, governing docs, any explicit skill→area map — into a manifest persisted at .speccy/<run-id>/capabilities.md; inject a phase-scoped "prefer these over generic approaches" block at every spawn site. Spec/plan research delegates discovery to project hunters before generic sweeps; breakdown attaches trigger-matched skills to each task and pre-resolves placement/existence questions (a build agent can invoke a Skill but cannot dispatch an Agent); the local-doc lens judges against the house skill set. Everything optional: no required config artifact, clean degradation to the generic pipeline, and routing needs no map — skills self-describe their triggers.

3. Prose readability pass (new)

Language-only rewrite of both skills' prose: point first in every section, one idea per sentence, long sentences split, parenthetical asides unfolded, condition sets as bullets. An independent fresh-context audit compared old vs new file by file — every rule, threshold, ordering, path, model assignment, gate, and exception survives; nothing lost, weakened, or invented. Code blocks (incl. the watchdog bash sketch, byte for byte), schemas, and frontmatter untouched.