Packaged source fidelity and the remaining C1 cost gap

中文镜像：同名检查点。

This checkpoint supports the single-session comparison. It qualifies source delivery and recovery on a bounded engineering workload; it does not establish a net efficiency gain or complete C1.

Versions and workload

The comparison pins original quota to mainline b201da8bab305d053866aa0d4516b97ba4a4499e and the proposed unified contract to 3afc2d547f6f9ca41368fd4f3f34d8f8dae9edcf in #7430. These are the tested versions, not a qualification of later mainline changes. Each isolated wheel has its own verified frontend source fingerprint; 1,857 packaged product files match that arm's source, excluding generated Chat assets, which pass their separate bundle check. Both use Python 3.14.

Each arm uses the same synthetic active Goal, full selected task and final stop, one current-Agent outcome and five peer outcomes with four requirements each. Both real File and SQLite authorities are initialized in disposable runtimes. The original 24 descriptions, owners, work links and global stop conditions are retained. No active Goal or installed host is used as a fixture.

The packaged CLI runs thin heartbeat bootstrap, one admitted App-heartbeat guard, saved-file reads, native lifecycle hooks, warm presentation, compaction recovery, refusal controls, writeback/spend and their replays. Hook events are synthetic; the actual packaged hook and CLI implementations execute. A deterministic observer reads full sources to check equivalence. That observer is not a model, a recommendation to read every peer body, or evidence of model retention. The actual emitted warm delta is separately read and reconstructed against its digest-verified baseline and current files.

This is a new matched workload, not a rerun or replacement of the historical 13,000-byte experiment. The earlier 25,862 > 13,000 failure is unchanged.

Source and recovery results

Both arms retain the selected task, all 24 requirements and global stops. Early spending is refused. Restoring the original capture recovers the same source without another admission; repeated writeback and spending commit one effect and one debit, leaving the original task open.

The mainline generic codec changes the admitted source's quota.compute from 1.0 to 1; its cold, warm-recovery and missing-baseline exact-source predicates fail on both backends. Its native cold presentation also carries 1 rather than the admitted 1.0. The C1 candidate preserves the original source through these paths. This exercises the source-preservation repairs also proposed separately in #7661 and #7657, rather than creating a caller-specific numeric policy. Those proposals and C1 adoption are distinct from a mainline or installed rollout.

Both native warm files use captured_file continuity. Their actual deltas reconstruct the current file exactly; missing baselines and damaged deltas recover its full body. This proves saved-file reconstruction, not retained model input. Mainline's reconstruction is exact relative to its already normalized presentation; that does not repair the preceding source loss.

A numeric-only capture mutation distinguishes two integrity boundaries. In C1, changing quota.compute from 1.0 to 1 still passes the standalone complete contract's JavaScript-normalized hash check. Reading the changed file against its original byte SHA-256 is refused; restoring the original bytes recovers it. The exact-byte capture boundary is qualified for this caller. The standalone contract's numeric commitment remains open; neither check substitutes for the other or qualifies arbitrary Host input.

Three preliminary author runs remain failed observations: the first fixture used a non-active Goal for refinement, the second exceeded the existing 512-character requirement limit, and the third supplied the forbidden fixed output-directory option to context-project. The corrected matched cohort uses an active synthetic Goal, bounded complete descriptions and the existing output-root option. No product limit, assertion or acceptance criterion changes.

Measured costs and decision

One observation per arm/backend gives the following byte counts:

Measurement	File: quota → C1	SQLite: quota → C1
Initial guard stdout (file manifest)	5,413 → 5,440	5,429 → 5,456
Cold saved presentation	65,756 → 68,411	65,866 → 68,519
Cold full-read CLI stdout	70,568 → 73,533	70,680 → 73,643
Actual warm delta file	3,761 → 3,755	3,769 → 3,763
Complete observer CLI stdout	482,505 → 494,464	483,299 → 495,250

The six-byte warm difference includes fixture-path and surface spelling; it does not demonstrate an encoding gain.

The observer makes 38 CLI calls in every arm, including fixture authoring, full recovery, refusals and readback. Separate warm-file qualification adds one CLI read per arm. In-process codec checks add nine coarse operations per arm; C1 also performs one extra standalone-hash negative control. These are explicit qualification costs, not a normal Turn's mandatory read policy.

The 38-call elapsed sums are 10.08 → 10.98 seconds for File and 9.41 → 9.57 seconds for SQLite. Sequential single samples, process/broker startup and the chosen full-read observer do not establish a latency effect. The warm follow-up timer includes CLI and codec verification and is not a CLI-latency measurement. Memory, model tokens/cache/billing, accepted task outcomes, installed App behavior and human-attention costs remain unmeasured. Construction and failed author probes also remain costs; they are not erased by the completed cohort.

The useful delta is source fidelity and a more complete contract, with modest additional cold/output bytes on this workload. There is no demonstrated net cost reduction. Continue through existing source/context/journal owners and independently usable repair PRs. Use actual caller failures to choose the next hash or bridge change; do not infer a need for another full-source carrier or parallel hash policy from an isolated assertion. Keep numeric integrity, caller/writer cutover, historical-reader retirement and the original-quota decision/outcome/full-cost comparison open. The optional two-session study remains separate.