Improve apply_patch one-shot matching

[yyjeqhc]

Summary

• replace model-facing strict_matching with explicit matching_mode={first_match,unique,exact_unique} and make unique the real default
• add bounded Exact -> TrimEnd -> Trim -> Normalized matching while requiring the final mutation target to be unique
• preserve stronger exact_unique stale-context/concurrency fencing and fail closed on legacy strict_matching at current Server ingress
• make EOF eligibility, repeated-context handling, later-chunk no-backtracking, ambiguity diagnostics, and Server-derived recovery deterministic and bounded
• add Runner capability migration, telemetry fields, and a deterministic matcher regression corpus

Safety / authority

• whole patch remains preflighted before execution; later ambiguity is zero-write
• SHA rechecks, rollback, project/session authority, PermissionEvaluator, stale/replacement Runner fencing, and outcome_unknown inspect-first behavior are unchanged
• ambiguous recovery is body-free and Server-validated; Runner cannot choose arbitrary recovery paths or preferred candidates
• old Server -> new Runner strict_matching remains wire-only compatibility; new Server rejects legacy strict_matching instead of silently weakening it

Validation

Fresh reviewer validation included:

• cargo test -p webcodex-core apply_patch — 25 passed
• deterministic matcher benchmark — 1 passed; 36 cases, ExactUnique 8/36 vs Unique 30/36, wrong-location writes 0, partial writes 0
• cargo test -p webcodex-runner file_apply_patch — 17 passed
• cargo test -p webcodex apply_patch — 23 passed
• focused tool-contract / ToolCall tests — passed
• cargo fmt --all -- --check — passed
• git diff --check — passed

Product-level Go/No-Go for keeping apply_patch as the default high-frequency model edit path should still be based on fresh sf/dogfood telemetry after deployment, not the deterministic corpus alone.