v0.7.7 — evidence-aware completion gate

Connects the v0.7.6 evidence model to the real agent loop. A turn-scoped,
harness-owned EvidenceCollector records changed files (from write/edit and
opaque mutating bash) and verification results (command, exit code, output
digest) straight from tool outcomes — the model cannot forge it. Every accepted
completion carries an EvidenceContract with verified; a failed done-gate emits
a structured completion_rejected record with its evidence, and a missing/absent
runner is captured as an explicit unverified assumption. Evidence resets each
user turn. The done-gate's accept/reject decision and the action schema are
unchanged — the contract is additive on assistant/completion_rejected records.