docs(workflow): remove the arbitrary repair-cycle cap

[TusanHomichi]

Primary Issue

Refs #73; supersedes the historical repair cap from #71 / #72. Depends on the separately reviewed fixture repair in #75 / #70.

Problem And Outcome

The workflow's two-cycle repair cap was invented during setup and the owner rejected it. Remove the default counter and mandatory waiver language. Continue evidence-driven repairs within the accepted scope until acceptance is met or a concrete blocker prevents progress; honor explicit owner budgets when supplied.

Changes

Update AGENTS.md, docs/workflow.md, CONTRIBUTING.md and the work-item template consistently, including delegation, resume and failed-check guidance. No product behavior, model route or required-gate change.

The test-only prerequisite #75 is merged as f154004. The final branch already includes it; the policy diff against current main contains only the four named documentation/template files and is byte-identical to the original independently reviewed policy diff. The prospective merge tree equals the tested candidate tree.

Verification

Frozen final candidate 65b207c69217cdf7e9fe61d3d63e2e787905131f, tree a10ea6ca5697626d7e3c9979f6e0385d6e13bd46.

PASS: composition review confirms the four policy files are identical to independently reviewed 1ec841a and retention.rs is identical to independently reviewed 0e2a21a; no merge conflict. Current documentation links/YAML/commands, whitespace and a fresh instruction-loading probe pass.

PASS: all nine local gates on this candidate, including 176 Rust tests and 10 browser tests. All 12 receipts (including documentation/instruction checks) have verified revision, clean state and actual log/artifact hashes.

PASS: hosted run 35885833462, 176 Rust tests and 10 browser tests. Final local receipts and composition proof were read back from the issue. All required gates passed; the original hosted failure remains archived.

Historical candidate 1ec841a passed 175 local Rust tests and 10 browser tests, but its hosted run failed with #70. Those receipts are preserved at their original revision; they are not current acceptance. The pinned SQLite/SQLx diagnostic and fixture repair are recorded in #70 / #75.

Review And Merge Notes

Luna/max drafted the documentation correction; the primary reviewed and corrected it, then an independent Astra/max model review found no substantive findings. Sol/max diagnosed and implemented the prerequisite; it also received independent model review. This does not claim human approval or per-turn runtime metadata unavailable from the host.

The existing setup merge/cleanup authorization covers this direct correction and its necessary fixture-only gate repair. No unrelated merge, deployment or product authority is inferred. The canonical graph and evidence remain in #73. Both prerequisites are verified: #75 is integrated and cleaned up, and every final-candidate gate passed. Merged as e0679bc133592eb465c0b707779ac84a1f98a1db; exact-main tree verified against the tested candidate. Evidence archive was read back and the owned branch/worktree and diagnostic scratch resources were removed. Unrelated work is preserved.