[Agent tooling] Make structured patches fail closed per file

[Chris0Jeky]

Context

The current overnight run hit the same atomic patch failure repeatedly: a multi-file apply_patch operation coupled unrelated exact-context anchors, so one stale or Unicode-heavy prose anchor rejected the whole edit. The run added an ad hoc ledger rule, but the failure recurred while fixing PR #1502. No tracked files were changed by either rejected patch, but each consumed review/verification time and made recovery harder to audit.

This now satisfies the two-strikes runway rule. The durable lesson belongs in the mirrored Taskdeck orchestration skills rather than another session-local reminder.

Scope boundary

Update only the active Codex/Claude orchestration skill guidance and its smallest validation/parity proof. Do not change product code, generic global laws, hooks, CI workflows, or unrelated skills.

Acceptance criteria

• The mirrored Taskdeck batch-orchestrator guidance requires one file (or one tightly homogeneous mechanical file set) per structured patch, ASCII-stable anchors for prose/Unicode documents, and an inspect-before-retry step after any rejected context.
• The guidance forbids retrying the same broad multi-file patch and requires recording an unresolved/repeated failure through the existing failure-capture route.
• Codex and Claude skill renderings remain behaviorally aligned.
• Existing skill validation/parity checks and docs governance pass.
• One bounded independent review and exact-head CI pass before merge.

Verification plan

Run the repository skill validators for both mirrors, the relevant parity/search assertion, docs governance, golden-principles governance, and git diff --check.