Family-aware template composition — frozen rc6 research

**Composition remains research only. Production receives a separate rc7 correctness guard; its template grammar and presets are unchanged.** Mode 6 in this fork is an experimental parameter representation requiring the research reader; the shipping reader must not be used for these archives.

Compare against the same frozen-source Native/Hybrid thin incumbent **including existing optional template factorization**. Template discovery and backend coding stay fixed. Count complete catalog, record-order IDs, template-to-column mapping, every parameter and column header, original-source SHA-256 and PIZD framing. Sparse-family pooling uses a fixed threshold of 16 records; no field names are hardcoded.

Each strategy below retains the current complete archive on a tie or loss. Minima within a strategy and across all candidates are exhaustive exploration, not a prefix selector or a free encoding policy. Grouping-only and codec-only gains overlap; do not add them.

Reference size positions

Reference sizes are the exact same public input hashes from the recorded rc6 panel. These are stored-size comparisons; reference timing is not reused for a speed claim. The exploratory after-column uses full evaluation and incumbent retention, not a production selector.

All source results and losing representations

Candidate-only minimum below excludes the incumbent, so regressions remain visible. The retained size includes it. Original archives and intermediate models remain under ignored `target/template-composition/archives/`; replay regenerates them.

Cost, correctness and limits

Full-archive candidates are decoded by the frozen research reader and compared byte for byte and by source SHA-256. Model reconstruction is also checked independently before backend encoding. Native payloads additionally verify exact intermediate-model reconstruction before framing.

Timing fields in results.json are one-shot **diagnostics**, with input I/O excluded. Model time builds all twelve variants together, including column codec search; it cannot establish production speed or per-strategy model costs. There is no speedup, RSS, large-stream, hostile-input-safety or universal-generalization claim. Family decoding currently materializes column values; a measured production decoder would need a separate design.

The thirty public excerpts are inspected development data, including tiny parser fixtures and four correlated GitHub windows. Six later windows are correlated continuations, not unseen systems. Fourteen samples are generated project data. The global grouping strategy may fragment sparse data and duplicate dictionaries; complete losses are retained in results.json.

Baseline source hashes, source archives and all input hashes are in inputs.json. Intermediate stopped-run artifacts are historical diagnostics, not passing final evidence. Current production may change independently during peer work; the report measures its frozen source only.

Reproduce

From the repository root with cached inputs: `python3 tools/template_composition_probe.py` prepares a fresh isolated project and refuses to overwrite an existing frozen project. Build with `CARGO_TARGET_DIR="$PWD/target" cargo build --release --manifest-path target/template-composition/project/Cargo.toml --example template_composition`. Run `./target/release/examples/template_composition docs/benchmarks/template_composition_rc6/inputs.json > docs/benchmarks/template_composition_rc6/results.json`, then this renderer. The exact research_sources.tar.gz archive is also retained for replay.

Instrumented search cost and possible pruning

One instrumented pass sums the following measured stages. These are not medians or a production latency comparison: concurrent verification activity shared the host, model construction generates twelve variants, Native packing includes intermediate-model self-checking, and all rejected candidates are encoded. Decode materialization is also research code. The search is expensive and these costs must not be presented as a fast preset.

A retrospective simplification retains all explored Native gains on all50 inspected inputs with just three candidates: pooled-family delta, pooled-family numeric codecs and positional numeric codecs. The same subset misses only75 B of Hybrid gains, on two generated inputs. This subset was chosen after viewing the data: it is a concrete implementation lead, not an independently validated selector or a measured speedup. A future implementation should construct only those representations and measure its actual end-to-end cost on additional sources.

Additional correctness evidence

The constructed controls add **260 complete archive/source SHA restores** across ten inputs and both backends: leading zeros, constant fields, fixed strides, mixed widths, CRLF, missing final newline, sparse-family threshold boundaries, binary anchors, 38-digit values and no-parameter templates. They are reconstruction controls, not public workload evidence.

The initial stopped NASA run exposed a separate production-code defect: literal-run lengths above65535 were truncated in the segregated channel plan. A minimal independent test reproduced it. The rc7 guard declines that plan, letting the existing flat coder preserve the data. All final measurements use this guard for incumbents and candidates. The original failing diagnostic is retained; no claim is made that this is an audit of every archive or hostile input.