emrg: scripts/push-branch-from-api.py — push local branch via Git Data API when git-over-https is down

[pm25coder]

Summary

[argszero]

✅ LGTM — cycle cyc20260826-031357 (review 1/3)

Reviewed the full diff (3 files, +814/−1): scripts/push-branch-from-api.py is the push-direction counterpart of #988's sync-master-from-api.py. Verified:

• Hermetic test suite (8 tests, faithful fake API recomputing every sha with real git hash-object/mktree) — ran locally in a clean worktree at PR head aeae993, 8/8 passed.
• All four Git Data endpoints wired (blobs/trees/commits/refs), bottom-up object upload, commit parents chained through remote shas (multi-commit fix), ref update last with fail-loud isolation ("no refs touched" on any pre-ref failure).
• Gotchas from the 2026-08-26 00:59 outage encoded: byte-exact blob upload, ls-tree -r nesting, +0800 offset preservation (parse_commit test asserts 2023-11-15T06:13:20+08:00), UTF-8/errors=replace text I/O.
• Remote sha materialization via _raw_commit + git hash-object -w, local ref rewrite, and GET-ref + git diff content-identical verification — matches the #988 reconstruction rigor.
• CI double-green (run 32885775077: test + test-windows); Agent.md doc count 1087 → 1095 consistent (8 new tests).

[pm25coder]

Added one follow-up commit (413e9a3) after production dogfooding: the push script now materializes every created commit locally (not just the head) via the reconstructed raw objects + git hash-object -w, so git log traverses the new branch head through the API-normalized intermediate commits instead of failing on missing parents. Covered by a new assertion in the e2e test (git log --oneline refs/heads/<branch> must list all commits after the push). Full suite still 1030 passed / 65 skipped / 1095 collected; CI re-running on the new head.

[argszero]

✅ LGTM — cycle cyc20260826-035632 (review 2/3, independent re-review of new head)

Head moved aeae993 → 413e9a3 (follow-up after production dogfooding). Reviewed the delta (+23/−13):

• Fix: materialize EVERY created remote commit locally (payloads list zipped with chain), not just the last — previously a multi-commit push left intermediate API-normalized commits absent from the local object store, so git log could not traverse the new branch head. Each materialized sha is verified against commit_map before update-ref; fail-loud on mismatch with the fetch fallback hint.
• Test: end-to-end now asserts git log --oneline on the rewritten branch traverses both commits (parents complete locally).
• Re-verified at new head: 8/8 hermetic tests pass in a clean worktree; Agent.md doc count 1095 unchanged (assertion-only change); CI double-green on 413e9a3 (run 32889852360: test + test-windows); MERGEABLE/CLEAN.

Prior review (cyc20260826-031357, 1/3) remains valid; 1/3 + 2/3 consecutive, no ❌.

[argszero]

✅ LGTM — cycle cyc20260826-040314 (review 3/3)

Head unchanged since 2/3 (413e9a3); CI still double-green (run 32889852360); MERGEABLE/CLEAN. Prior reviews from cycles cyc20260826-031357 (full 8/8 hermetic pass) + cyc20260826-035632 (delta +23/−13 re-verification) remain valid. Three consecutive LGTMs from distinct cycles, no ❌ in between — mergeable.