Skip to content

bench: add a prepared literal-range patch reference - #96

Draft
gspark-etri wants to merge 5 commits into
exp/88-timing-evidencefrom
exp/95-literal-range-patch
Draft

gspark-etri wants to merge 5 commits into
exp/88-timing-evidencefrom
exp/95-literal-range-patch

Conversation

@gspark-etri

@gspark-etri gspark-etri commented Nov 2, 2026 •

Copy link
Copy Markdown
Contributor

Compare admitted in-place materialization with a sound prepared literal patch under a common file-readiness contract. The reference already preserves same-offset bytes; the experiment tests the actual executor tradeoff instead of attributing that operation uniquely to the materializer.

The stdlib reference and public legacy-diff adapter bind input/diff/keep/source/manifest identities and verify kept bytes. The new cohort freezes input/store/cache/patch/driver/build-receipt pins, uses private same-filesystem donor copies and charges preparation. R verifies the full base, invokes the actual admitted driver and syncs/full-hashes its private output; L runs the actual patch CLI with its corresponding checks. Both require donor-inode transfer and exclusively publish the verified target. A separate full-SHA audit follows the timed endpoint. Actual initialization and validation costs remain charged. External build/host attestation is still required.

An excluded RL smoke gates five paired blocks RL/LR/RL/LR/RL. Main rechecks the pinned smoke's actual method, identity, clock and inode witnesses. A failed attempt stops the matrix without retry or valid-subset aggregate. Success-data cleanup is frozen, runs only after the audit, and removes guarded private child directories while retaining raw logs/results. Existing origin/cache layout is admitted read-only before any driver call, preventing a mistyped immutable input directory from gaining backend directories. Model mode accepts only the preregistered complete directed decode pair on Linux; fixtures cannot be relabeled as model evidence.

Validation: head 7726500, base e1b29c7, tested merge 8b134aa. CI41556702186 passes all three jobs; complete decoded logs were retained and checked. Go fmt/vet/race/build,77 existing CPU/evidence checks,39 reference/cohort checks and both actual S3 suites/roundtrips pass. The Go integration actually performs R/L smoke and all five synthetic paired blocks with full target audits. Local Windows:37pass+2explicit Go integration skips. Important origin-mutation feedback has observed RED→GREEN regression and independent recheck; no remaining Important issue in the bounded review.

Earlier CI41551086659's unchanged causal trace-parser failure remains retained and unresolved. No parser/core change or retry-to-green fix is claimed. The separate retained synthetic diagnostic traces are new evidence and do not replace that earlier missing trace.

Benchmark only: no node/controller call, accelerator allocation, installation or production/runtime modification. Prepared raw literals are not codec-matched storage controls. Origin data and earlier measured sources remain unchanged. Real-model R/L timing: not_run. Keep draft/unmerged until actual admitted smoke/main and raw replay are available.

Refs #95.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant