[yifanxuaaa]

Summary

Follow-up to #472 / #473. The payload chain-work budgets were fixed at 512 KiB canonical and 256 KiB encoded, while the construction cutoff can be raised to 1 MiB. Before #473, a whole file above those sizes failed the next Save. After #473 it is stored FULL, so a Store with a raised cutoff gets no whole-file deltas for files of about 256 KiB or more: every version is a complete copy.

• The budgets are now 4 × cutoff canonical and 2 × cutoff encoded (StorageCapacities::from_policy). At the default 128 KiB cutoff that is exactly 512 KiB / 256 KiB, so default Stores do not change.
• At every supported cutoff, the largest whole file fits a chain alone, so the producer never stores a base that a read could not charge, and large whole files keep deltas.
• Trade-off (owner-approved): one read may reconstruct up to 4 × cutoff, which is 4 MiB at the 1 MiB cutoff. That is the same ratio to the largest whole file that the default Store already accepts. A depth never raises the budgets. Chunk chains (32 KiB chunks) stay depth-bound and are unaffected, so chunked files of any size behave as before.
• fix(storage): store FULL when a delta base alone exceeds the chain budget #473's early refusal stays as a guard. Valid output under a Store's own policy no longer reaches it.
• No schema or format change. Components must be version-matched (release policy). Upgrade note: existing Stores with a raised cutoff start producing whole-file deltas after upgrade. Their existing chains already fit the larger budgets.

Tests

• layerfs-storage/tests/chain_budget_scaling.rs (new): derived budgets at every cutoff, with the default unchanged; the largest-whole-file-fits-alone invariant; Save fails with Integrity("dependency canonical work" | "dependency encoded work") when the delta base alone exceeds the chain budget, instead of storing FULL #472's shapes now chain at depth 1; the largest file at each cutoff; a 600 KB chain restarting at the scaled 4 MiB budget with exact readback.
• delta_base_budget.rs (from fix(storage): store FULL when a delta base alone exceeds the chain budget #473):
  • Still-valid cases move to the default cutoff, where the same numbers hold: the exact canonical and encoded fills, same-Save vs separate-Save parity, and the corrupt-base check.
  • The early refusal is reached by lowering the test Store's cutoff after the base is stored.
  • Its five 1 MiB "base alone over budget" cases are deleted. Under scaled budgets those inputs now chain, which chain_budget_scaling.rs covers.
• sdk_large_file_budget.rs (from fix(storage): store FULL when a delta base alone exceeds the chain budget #473): now asserts a whole-file delta for fs.apply, fs.write and a Workspace Commit. It keeps Init and the default-cutoff control.

Side finding (not changed here)

At zstd level 3 with no long-distance matching, highly repetitive text (word salad from eight words) loses the delta to FULL even for a 10-byte edit: the matcher takes nearby repeats over the long match into the base. CSV-like text deltas normally, and the SDK delta tests use that. This is codec behavior and independent of the budgets.

Test plan

• Fail-before against main 924f782 production code: chain_budget_scaling 1 passed / 5 failed; sdk_large_file_budget 2 passed / 3 failed (work_exceeded, FULL, no delta). With the fix, all pass.
• layerfs-storage 93 passed; layerfs-persistence 118 passed, 2 ignored; layerfs-project 23; layerfs-store-service 6.
• Fast-lane selection (tools/ci_tests.py) per package: layerfs-api 45 binaries, 138 passed, 6 ignored; layerfs-daemon 9 binaries, 24 passed, 1 ignored. Run package by package because cargo-nextest is not installed locally and the worktree launcher refuses workspace aggregates.
• Clippy -D warnings (storage, api, all targets), cargo fmt --check, tools/check_product_boundary.py, 155 tool tests.
• Not run: Linux, mounted FUSE Commit, Windows, read latency/memory at the deepest 4 MiB chain. Durable: NOT_RUN — disabled by owner until explicit reauthorization.

All runs were on macOS arm64, Rust 1.85.1, through tools/worktree.py private resources, with Disposable Stores only.

Production LOC: 102,580 -> 102,588 (delta +8). core 38,387 -> 38,395; layerfs-storage 9,060 -> 9,068 (store/policy.rs 179 -> 187); engine 61,534 and api 2,659 unchanged. Reference absent (retired R9c; last counted 65,417 at tree 498dd1917812). Method: tools/production_loc.py (blob 3341e4d036) --detail/--files --root on a git archive of 924f782 and on this tree.

Generated with Devin