feat(server): integrate Riftri worktree storage

[Kinfe123]

Problem

T3 creates ordinary Git linked worktrees, which duplicate checked-out file storage. This draft adds a native copy-on-write storage path while retaining real Git worktrees and recovery safeguards.

Integration

• Bundle released Riftri 0.6.4 in desktop and standalone CLI artifacts; no global Git replacement or user installation is required.
• Route supported worktree lifecycle operations through the storage adapter. The fork defaults to auto mode; explicit Git mode remains available. Auto fallback is limited to pre-mutation policy refusals; operational failures remain errors.
• Recover managed storage at startup and guard file, terminal, and provider access when recovery fails.
• Cover lifecycle behavior, private-write isolation, allocation, packaging, and creation latency in CI.

Fork-only scope

This remains a draft in farming-labs/t3code, not a submission to pingdotgg/t3code. Fork PR #2 is unchanged. Any upstream proposal should first seek maintainer agreement on an opt-in scope; this fork's auto default is not a proposed upstream default.

The base verify/upstream-a6ec88f7 contains upstream a6ec88f7a716fc421bd22c2484881c44110f9375 plus a separate hosted-runner substitution commit. Those runner changes are shared by base and head and are not in this PR diff. The integration commits can be applied independently; the fork CI base is not intended for upstream. This upgrade does not rebase onto later upstream changes.

Verification — Riftri 0.6.4

• All 19 CI jobs passed on upgrade commit b4c3d65be933751c99dff232d95136c4256fe64d: complete run and measurement artifacts. This includes all six server shards, web and other workspace tests, typecheck, lint, native APFS/OverlayFS storage tests, the complete desktop archive's bundled native runtime, and standalone CLI startup.

• The upgrade commit changes only the dependency pin, nine-package lock graph, exact-version release-age exceptions, and dependency test fixture.

• All nine package integrity hashes match the npm registry; provenance receipts were verified.

• Local targeted formatting, frozen-lockfile validation, 20 CLI external-dependency tests, and the isolated bundled Node adapter test passed.

• Broader local suites could not load new upstream dependencies in the lightweight local dependency fixture. They are not claimed as passing locally; CI uses the complete dependency installation.

Measured storage and creation time

The real GitVcsDriverCore fixture creates three worktrees of the same tree on an isolated volume: 32 MiB incompressible data plus 257 text files. Allocation includes the initial immutable base and all three views, measured with filesystem free-space deltas rather than summed du.

Backend	Git allocation	Riftri allocation	Saving	Git create	Riftri cold create	Riftri cached create
macOS APFS	107.47 MB	36.22 MB	66.30%	128–369 ms	735 ms	447–519 ms
Linux OverlayFS	107.22 MB	36.04 MB	66.39%	67–87 ms	357 ms	194–286 ms

MB above is decimal. Reports are attached to the current CI run as riftri-performance-macos-15 and riftri-performance-ubuntu-24.04, and identify riftri 0.6.4.

Creation is slower than Git in this fixture. Each backend has one cold and two cached observations, not a latency distribution or evidence of a speedup over 0.6.3. Dependencies, ignored files, and build outputs are not cloned or shared. These results demonstrate tracked-checkout storage savings, not a 66% reduction in total user workspace storage.

The fixture also verifies clean Git status, matching HEAD, exact-base reuse, private-write isolation, and managed cleanup. Integrity checks, journals, and recovery are not skipped to improve timing.

Model: GPT-5. Harness: Codex.