Historical snapshot: the canonical repository body is retained because a harness consumes it. Keep edits to that versioned input in the repository; this wiki copy records its migration snapshot.

Historical report imported 2026-10-10 from docs/research/cross-system-source-synthesis.md. Original SHA-256: 1aa6e2748af7f23acc6151e895602afb72d37afd08aabcdc7231c80427ba50dd. Measurements and outcomes retain their stated revisions. Links and local environment labels are normalized; $FABRIC_DATA is the mounted experiment drive. See evidence policy.

Source mechanisms to test in Fabric

Research synthesis, 2026-10-08. The owner requested downloading and dissecting the supplied FOSS sources, then using their engineering to generate and test new trajectories. This is implementation research, not a storage migration or an upstream benchmark. The finite continuation (local checkout reference docs/experiments/benchmarks/cross-system-continuation-protocol.md; not present at origin/main) registers containment, provenance, workloads and evidence limits.

What was downloaded and inspected

Twenty-four additional repositories now have complete regular-member source inventories. The matched sweep (local checkout reference docs/experiments/benchmarks/cross-system-sweep-protocol.md; not present at origin/main) repaired the three earlier incomplete downloads: ClickHouse at its existing pin, and Vector/FoundationDB at the explicit replacement revisions already used for selected-file inspection. ClickHouse required a 512 MiB compressed-archive allowance and a 16 MiB compressed-inventory allowance. Its complete 361264396-byte archive was reused exactly after the first inventory attempt failed; it was not downloaded twice. The historical 128 MiB cap and HTTP 404 failures remain recorded. New pins do not turn the historical revisions into successful retrievals. Turso's earlier archive and 24 matching reviewed blobs remain separate evidence. The Feldera storage-design proposal is counted separately from its implementation repository. Complete retrieval does not imply a recursive submodule checkout.

Selected implementation files, licenses, immutable revisions, complete archive member inventories where available, requested-path failures, archive hashes and cleanup receipts are retained under cross-system-run-01/source (local evidence omitted from stage). The earlier full successful archives were removed after verified inventory and selected-source preservation. The three repaired complete archives remain under cross-system-sweep-01/source (local evidence omitted from stage). The new independent checker reread every regular member of the three repaired archives and bound each selected excerpt to its full inventory. Its self-consistent excerpt-substitution control catches a provenance gap that separate archive and excerpt checks missed. The bounded historical failed ClickHouse prefix remains. Retrieval executed no upstream application or build script; later registered workload execution is separate from this download claim.

Dissection	Systems	Mechanisms actually inspected
Storage	Comet, Feldera implementation and proposal, Vortex, Lance, redb, Fjall	Reservation ownership; cancellation drain; retained versus construction memory; merge priority and fuel; compressed predicates; snapshot owners; oldest journal obstruction
Query	Tantivy, Quickwit, DuckDB, GreptimeDB, Tempo, Roaring, ClickHouse	Candidate IDs versus payloads; index maintenance; sound early termination; serialization overlap; trace-specific lookup; container dispatch; saturated-index fallback
Operations	Vector, OTLP Arrow receiver, FoundationDB, Shuttle, Loom, fail-rs, proptest, Prometheus, SlateDB, differential-dataflow	Admission lifetime; decode before reservation; modeled persistence loss; exploration stop reasons; failpoint activation; replay/shrinking; numerical codecs; publication/reclamation; input versus output work

These are selected-function dissections. An archive inventory does not imply a whole-engine audit, transitive dependency review, license opinion or measured performance advantage. Exact paths, source links and transfer limits are in the three dissection records.

Patterns that survived comparison

Account for the lifetime that consumes the resource. Comet's explicit drain and Fabric's existing completion-owned query permit address the same ownership boundary. Lance constructs ranges before deciding whether to retain them; OTLP Arrow converts input before queue-byte admission. Those are concrete counterexamples to equating a retained cache or queue allowance with the whole process's peak memory. Fabric should observe construction, retained owners, worker completion, allocator state and cgroup memory separately. Independent peak measurements cannot be added to reconstruct simultaneous pressure.

Relieve the oldest real obstruction. Fjall's journal manager inspects what prevents reclaiming the oldest journal; Feldera also considers batch fanout before resident memory. Fabric's ordered-prefix repair removes one unnecessary wait, but checkpoint work still shares the commit thread with ingest. The experiment must measure both retained journal byte-time and ACK tails.

An acceleration structure needs permission to stop helping. ClickHouse's saturated set becomes non-pruning. DuckDB disables a cardinality shortcut when filters invalidate its premise. Tempo's trace lookup depends on ID ordering; its general search takes another path. Fabric should preserve exact fallback and choose a candidate representation from predicate semantics and measured shape, rather than mandate one index for every field. Tokenization does not implement arbitrary literal substring matching.

Budget consumed work, including work with no output. Differential-dataflow's inspected merge can consolidate many inputs to zero while its output-based fuel charge remains unchanged; the Spine also supplies fuel independently to layers. Fabric should measure input bytes, decoded rows and comparisons before adopting a scheduling quantum. Weighted consolidation is not custody deduplication.

Model the state that survives a failure. FoundationDB distinguishes pending writes, sync and simulated persistence loss. Vector distinguishes reader progress, durable checkpoints and asynchronous physical deletion. A successful function return, visibility to a local reader and durable custody are separate events. Fabric's recovery experiments should preserve requested/applied fault cuts and resulting file images, not count every injected error as equivalent coverage. This does not establish real-device power-loss behavior.

Derived models and competing trajectories

These are deductions and hypotheses, not calibrated controllers or novelty claims. They refine the existing D1–D5 cards rather than add unlimited prototypes.

For maintenance, measure J = integral(retained_journal_bytes(t), dt) together with ACK latency and live memory. A candidate action's useful relief is the eligible prefix it actually releases, not the size of its own output. Compare oldest-obstruction, largest-resident and FIFO policies independently before combining them. The null is that checkpoint contention or delayed input reclamation cancels the apparent benefit.

For optional indexes, let C_build include construction, publication and future maintenance; let delta_C_query be the measured saving per applicable query. Only when delta_C_query > 0 is the simple amortization threshold q_break_even = C_build / delta_C_query meaningful. Expiry, repeated conversions, snapshot ownership and memory occupancy are additional constraints. Test zero, one and repeated queries, clustered and shuffled rows, and common predicates. The null is that reuse never pays for the extra representation before expiry.

For bounded scheduling, a target quantum Q requires counting consumed work. Even with correct accounting, work can exceed Q by the largest indivisible operation; there is no fixed latency bound without bounding that operation and its cost. Measure cancellation-to-worker-completion and oldest-prefix progress alongside throughput. The null is that more yielding retains inputs longer and increases total byte-time without improving responsiveness.

For compressed predicates, Vortex gives a concrete mechanism, not a general permission to skip decoding. Charge conversion, dictionary construction, candidate verification, exact literal fallback and final materialization. Compare with Fabric's existing Walk and integrity checks on the same output. The null is that preparation and verification dominate the saved payload work.

The census and matched workload sweep have now exercised these mechanisms. The sweep includes native sort-cap and phase-attribution measurements, Arrow/DuckDB/Vortex query comparisons, RowSet representation/locality controls, prepared scan reuse, and real Store admission cadence. Its research vectors incorporate adverse regimes and rejected predictions. Nomination remains constrained by measured costs and the existing two-prototype ceiling. No production format, durability ordering, query semantics or dependency changed in this source-research round.