Changelog: pushing no longer fails on non-linear local history (6430105805)

[DingDingChae]

Pushed 5504d68614..6430105805 to main, carrying fix commit fa6f9a954b plus a merge that preserves the user's own 11af4be095 AGENTS.md edit. CI on the pushed commit: see the run list for 6430105805 — not yet verified at publication time.

Fixed

• Push failed outright with "The local-only history is not linear." (Refs Push fails outright with "The local-only history is not linear" / 有merge commit就推唔到 #58). Legacy local-commit batching runs before pushRepo at every push call site, and its read-only preparation throws when the local-only range is not a straight parent chain. A merge commit makes it non-linear — completely ordinary, and something Git pushes without complaint — but the refusal propagated and aborted the entire push, so nothing was pushed at all.

  Batching is an optional repair for histories written by older app versions. A LocalCommitBatchingGitError raised by the preparation phase is now logged and absorbed so the ordinary push proceeds. This is safe precisely because that refusal is raised before any mutation; a failure once rewriting has begun still propagates, so a half-rewritten history is never pushed blindly. A source-guard test pins both halves: the catch/return shape, and that the mutating call sits outside the guard.

Local verification

Root npx tsc --noEmit exit 0; prettier clean on both touched files; batching and guard suites 44/44 across five files (local-commit-batching entry points, pending-commit-push safety, cheap-lfs commit entry points, commit-push-batching, submodule mutation guard).

Related investigations filed

• Committing more files to a Cheap LFS repo re-downloads pinned payloads / 加檔案commit會重新下載大檔 #55 — committing re-downloads Cheap LFS payloads. Root-caused: after a successful pin no copy of the payload remains on the machine (original deleted, quarantine removed, temp upload copy deleted in finally), and materializePointer has no local-source fallback. Fix queued behind the in-flight hashing work due to a hard coupling.
• Cheap LFS batch upload: spurious 'reviewed release/asset changed' errors mid-batch / 批量上傳無故話release變咗 #56 — spurious mid-batch "reviewed release changed" errors. Root-caused: the append guard compares the whole release fingerprint including siblings' label, updatedAt and downloadCount. Fix in progress.
• Commit-and-push proofs run 'git add -A' into a temp index, re-hashing multi-GB Cheap LFS payloads / 證明步驟走去hash晒啲大檔 #59 — new: commit-and-push proofs run git add -A into a temp index, hashing multi-GB Cheap LFS payloads into .git/objects.

---

推咗一個好緊要嘅修復：以前未推嘅歷史入面有個merge commit，成個push就直接死俾你睇，紅字話你「歷史唔夠直」。但Git本身推merge commit推得好開心㗎!嗰個「執靚舊commit」功能純粹係加分項,做唔到就應該靜靜哋讓路,唔係賴死喺度阻住人推嘢。而家淨係「郁手之前」嗌停嗰啲當冇事照推;真係改到一半先仆街嘅,照舊即刻收手,唔會推半截歷史累街坊。