Historical report imported 2026-10-10 from docs/experiments/benchmarks/catalog-mixed-bounds-findings.md. Original SHA-256: e1617533d578c68306039d98a47ac9ee6e2574c71c9be76b91fe55d9c5d427ef. Measurements and outcomes retain their stated revisions. Links and local environment labels are normalized; $FABRIC_DATA is the mounted experiment drive. See evidence policy.

Mixed-source and numeric-bound findings

Registered additions (local checkout reference docs/experiments/benchmarks/catalog-coverage-interleavings-protocol.md; not present at origin/main) ran through the serial resource coordinator. Results belong to their captured source/diff and protocol hashes, not later source changes. No new workload ran for this note.

Mixed sources and preserved fixture failure

catalog-mixed-sources-01 failed, Cargo exit 101, job 11.546362403 seconds. The fixture required directory labels [1,3], but enumeration returned [3,1]. The contract requires membership, not directory ordering; sorting the fixture's local labels fixes that assumption without changing production. This original failure stays failed. Its 60 successful oracle calls and store/counterexample remain under owned data-drive evidence: $FABRIC_DATA/evidence/fabric-work-af68cf661d834efb824302b338c05ef2. The launcher relocated original scratch; root's later cleanup receipt (local evidence omitted from stage) records byte-verified archival and removal of only this failed owned root. The archive (local evidence omitted from stage) preserves the failed outcome; 4,223,853 source bytes were removed after verification.

Fresh catalog-mixed-sources-02 passed, Cargo exit 0, job 14.637652678 seconds. Receipt (local evidence omitted from stage) and oracle artifacts (local evidence omitted from stage) support 120 calls with expected pass, exit 0, passed=true, no violations: 15 per plan at each fresh/reused before/after reclaim population. The actual fixture had published Segment 1 with its unreclaimed journal, pending journal 3, and active groups 5–8. Real Store/sealer published 3 and reclaimed both sealed journals while active data remained queryable. The passing Rust assertion compares exact raw replay bytes/identities/multiplicity against the independent pre-IO ledger; the replay output is not separately archived. This is a sequential coverage replacement test, not a concurrent reclaim schedule.

Whole-job peak 538,693,632 bytes, sampled scratch 8,262,262 bytes, archived evidence 8,326,333 bytes; successful scratch removed. Historical failed job peak was 1,031,053,312 bytes, not a product memory regression.

Numeric statistics and negative controls

catalog-numeric-bounds-01 passed, Cargo exit 0, job 1.003439025 seconds. Receipt (local evidence omitted from stage) and artifacts (local evidence omitted from stage) contain exactly two accepted absent-statistics calls (Scan/Walk) and four rejected negative controls (reversed or plausible false-negative bounds, both plans). Every rejection has exit 1 and passed=false; together they record 576 ROW-DROPPED violations. Table data pages remain unchanged while time-column footer statistics and manifest size/digest are rewritten. Missing statistics produce conservative bounds and exact answers. Authentication does not protect against semantically lying numeric bounds: those answers were intentionally wrong and detected independently. No runtime protection claim follows.

Whole-job peak 73,719,808 bytes, sampled scratch 308,109 bytes, archived evidence 386,924 bytes; successful scratch removed. Both successful receipts record zero OOM events, zero swap and no descendants requiring termination. Resource numbers include graders/cache; scratch sampling may miss transients. Partial-table oracle semantics, general concurrent schedules, typed-handle IO leases and qualification remain outside these results.