Turn off git auto-maintenance in the archive

[PEEKPerformer]

What this PR does

Sets maintenance.auto=false and gc.auto=0 in the archive's local git config, so a holotype commit never starts a background repack. init.py sets them for new archives. ingest.py sets them for existing archives on their next tick.

Why

git runs git maintenance run --auto --detach after every commit. When the loose-object count passes gc.auto, that starts a detached full repack (git repack --cruft and git pack-objects --all). A second commit starts a second repack even while the first is still running. Ingest makes one commit per changed session, so a busy tick starts several at once.

An archive's pack is mostly zstd or git-crypt blobs, which do not compress or delta. Each repack reads the whole pack and holds gigabytes. On a real archive (7.3 GiB pack, 16 GB Mac):

• One ingest tick left 63 git processes holding about 44 GB, alive for more than 12 hours.
• macOS logged repeated low-memory kill events, and the machine had to be force-restarted.
• The repacks never finished. The pack file was last written two weeks earlier, so the loose-object count stayed over the threshold and every new commit started another repack.

Reproduced on git 2.54 with a scratch repository: four commits in a row left five concurrent pack-objects --all processes. With the two settings in place, the same commits started none.

Turning auto-maintenance off is safer than tuning it. Tuning would still let repacks overlap, and repacking incompressible blobs gains nothing. Loose objects now wait for a manual git repack -d, which packs only the loose objects and leaves the existing large pack alone.

Type of change

• Bug fix (non-breaking change that fixes a documented or observed defect)
• New feature (non-breaking change that adds capability)
• Breaking change (alters existing behavior; requires major-version bump and CHANGELOG entry)
• Documentation only
• Performance improvement (no functional change)
• Internal refactor (no user-visible change)
• CI / tooling

Non-negotiables check

• Forensic completeness (no filtering of transcript content)
• Provenance via hash chain (SHA-256 per session, verifiable with stock Unix tools)
• Git is the archive substrate
• Environment capture (host CLI version, OS, holotype version, project-repo git state)
• Append-only by design
• No silent network behavior

The change touches only .git/config in the archive. No tracked content changes.

Selftest

• python scripts/selftest.py passes locally
• If this PR adds a feature or fixes a bug, a new selftest assertion exercises the change
• CI is green (GitHub Actions workflow on this PR)

Two new steps: init.py sets both keys, and ingest.py restores them on an archive where they were unset.

Manifest / archive-format impact

• No archive-format change
• manifest_version bumped from N to N+1
• index_schema_version bumped from N to N+1
• On-disk archive layout changed (describe migration path)

Documentation

• CHANGELOG.md updated (under the next-release section, 2.4.1)
• README.md updated if user-facing behavior changed
• SKILL.md updated if wizard flow changed
• docs/ updated as appropriate

Cross-platform impact

• Verified on macOS
• Verified on Linux (CI, ubuntu-latest)
• Verified on Windows (not applicable: no Windows runner configured)
• CI passes on all configured runners

AI assistance

Diagnosis, fix, and draft written by Claude Code from the macOS memory-kill reports and a scratch-repository reproduction. Selftest run locally.

Anything else reviewers should know

Out of scope: run_git in ingest.py still has no timeout, so a hung git push can block a tick. That did not cause this incident.