Speed up checkpoints v2 pre-push recovery

[computermode]

https://entire.io/gh/entireio/cli/trails/374

This PR improves the performance of the pre-push hook for v2-enabled repos with 100+ checkpoints down from ~60s (!) to 5s. Tested with the entiredb repo. Observed push times went down from close to a minute to about 5-6 seconds (still unacceptable but much better while the refs fixes are being addressed). Double-checked and ensured checkpoints were pushed up correctly.

• Replace the v1 <> v2 missing checkpoints check for the migration hint message with a single "does v2 /main exist?" ref probe. The hint now only fires when migration was never run; partial migrations are no longer flagged (a little drift is acceptable post-migration since we fall back to v1 in the UI when we display transcripts for now anyway).

• Probe rotation archive refs with --filter=blob:none and top up only the matched archive's blobs after pick. Falls back to a full fetch on any first-fetch error (e.g. server without uploadpack.allowFilter).

• Bind the per-archive ancestry walk to 1s. On repos whose archives are fully disjoint from local /full/current we'd otherwise burn seconds concluding nothing matches. A future /full/root anchor will replace this with a constant-time lookup ( see v2 checkpoints: add deterministic /full/root anchor commit #1201).

---

Note

Medium Risk
Changes the v2 pre-push recovery fetch/rotation-conflict logic to use explicit filtered fetches and a new archive-selection path, which could affect correctness in edge cases (servers without filter support, unusual ref layouts) but is scoped to checkpoint sync/push flows.

Overview
Speeds up v2 pre-push recovery by switching recovery fetches to probe with --filter=blob:none (with fallback to unfiltered fetch) so merges can operate on commits/trees without downloading transcript blobs.

Adds remote.ResolveFilteredFetchTarget to always resolve remote names to URLs for explicit filtered fetches, preventing git from persisting promisor/partial-clone settings onto named remotes.

Refactors rotation-conflict handling for v2/full/current to fetch archived generations via a wildcard refspec, select the related archive using git merge-base with a bounded per-archive walk budget, and improves temp-ref cleanup; updates/adds tests to assert no blob downloads and no remote.origin.promisor/remote.origin.partialclonefilter side effects.

^{Reviewed by Cursor Bugbot for commit 1bc7c4e. Configure here.}

[Copilot]

Pull request overview

This PR optimizes v2 checkpoint pre-push recovery by reducing expensive migration checks and limiting how much checkpoint archive data is fetched during rotation recovery.

Changes:

• Replaces v1/v2 checkpoint diffing for the migration hint with a local v2 /main ref existence check.
• Changes rotation archive probing to use blobless fetches, then fetch blobs only for the matched archive.
• Adds a wall-clock budget for archive ancestry matching and tests for blobless archive probing behavior.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 2 comments.

File	Description
cmd/entire/cli/strategy/push_v2.go	Updates remote rotation archive fetch/recovery to use blobless probing, blob top-up, and ancestry walk budgeting.
cmd/entire/cli/strategy/push_v2_test.go	Adds coverage for avoiding blob fetches from unmatched rotation archives.
cmd/entire/cli/strategy/push_common.go	Simplifies v2 migration hint detection to probe for the local v2 /main ref.
cmd/entire/cli/strategy/push_common_test.go	Updates migration hint tests for the new v2 /main existence behavior.

[computermode]

Bugbot run

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

Comment @cursor review or bugbot run to trigger another review on this PR

^{Reviewed by Cursor Bugbot for commit 1bc7c4e. Configure here.}

[cursor]

Merge-base exit codes not distinguished, masking errors

Medium Severity

commitsShareHistory uses cmd.Run() == nil which treats every non-zero exit code from git merge-base identically. Exit code 1 means "no common ancestor" (a valid negative result), while exit code 128 signals an actual failure (bad object, corrupt repo, etc.). By collapsing both into false, a real git error during the ancestry walk is silently swallowed, potentially causing the rotation-conflict recovery to skip the correct archive and produce a misleading "no remote archive shares history" error.

 

^{Triggered by learned rule: Distinguish git command exit codes — exit 1 means "no results", not error}

^{Reviewed by Cursor Bugbot for commit 1bc7c4e. Configure here.}