Add multi-outcome factor-model input to the trellis engine

[ciaranra]

Summary

Adds an N-ary (multi-outcome) factor-model input to exp/pecos-trellis, closing the largest capability gap versus the upstream Frontier decoder's choice engine: a factor with arbitrarily many mutually exclusive outcomes, each carrying its own probability, detector toggles, and observable toggles. This also provides the public model-construction API (FactorModel / Factor / Outcome) that previously existed only upstream.

• FactorModel::new validates outcomes (probabilities in range and summing to 1 within 1e-10, indices in range, no duplicates per outcome); TryFrom<&SparseDem> maps each binary mechanism to a two-outcome factor.
• TrellisDecoder::from_factor_model classifies the model: binary-shaped models delegate to from_sparse_dem and are bitwise-identical to the equivalent DEM build (BP scoring and indistinguishable-merging fully supported); genuinely multi-outcome models run a new N-ary DP kernel.
• The N-ary kernel mirrors the binary DP (same merge, prune, ordering, and forced-layer semantics) with per-outcome log-prior branches, and generalizes suffix-compatibility scoring to per-row toggle probabilities.
• A two-outcome factor delegates only when its stored baseline agrees with the computed complement to within 1e-12 relative — float-roundoff-level agreement in either listing order; larger disagreement is deliberate parameterization and is evaluated faithfully on the N-ary kernel with each outcome's stored probability.
• Binary-only features fail loud on genuinely N-ary models: BP-guided pruning is rejected whenever it would engage (inert on the unpruned fast path, matching the binary kernel), and indistinguishable-mechanism merging is rejected unconditionally.
• deadline_column_order_for_factors / backward_deadline_column_order_for_factors order factors by the union of their outcomes' detector supports.

Binary path unchanged

The binary kernel is under a byte-identical external parity contract. Every existing fixture (upstream_fixtures.json, upstream_order_fixtures.json, bitwise_snapshot.json) passes unmodified; the refactors on the shared path (kernel enum wrapping, mask-slice plumbing in merge_branch, integer-only ordering refactor) move zero float operations, verified by an independent mechanical diff of the decode functions.

Verification

• New frozen oracle fixtures (upstream_nary_fixtures.json): 7 N-ary models decoded per-shot against the upstream choice engine, cross-checked against the upstream pure-Python reference on unpruned decodes, hand-verified on the degeneracy case. PECOS matches status, predicted label, every terminal log mass, and log evidence within 1e-9, unpruned and pruned (pruned rows also pin the retained-label set, validating the generalized suffix scoring).
• Brute-force enumeration oracle over seeded random N-ary models, all syndromes, exact-decode regime.
• Equivalence: two-outcome factors with nonempty baselines vs their forced-toggle binary DEM image — identical decisions, masses within 1e-9.
• Delegation: DEM-derived and hand-written decimal-literal factor models decode bitwise-identically to the DEM build, in both outcome listing orders, including with BP scoring and merging enabled.
• Every new guard (probability sum, empty factor, index validation, delegation gate both directions, BP and merge rejection, column-order permutation) is mutation-tested.
• cargo test -p pecos-trellis -p pecos-frontier -p pecos-bp-trellis, cold cargo clippy --all-targets -- -D warnings, and cargo fmt --check all pass.

Follow-ups (separate PRs): integer max-log metric mode, Python bindings and user-facing docs.

[ciaranra]

Review-panel follow-up pushed: the binary-delegation gate now accepts pairs whose induced baseline log-mass error is within the engine's 1e-9 acceptance bar (the previous 1e-12-relative bound was unreachable for small baselines, demoting exact-complement factors like (1e-6, 0.999999) off the binary kernel). New tests: the repro in both listing orders (bitwise vs the DEM build, BP enabled), a tiny-baseline case documenting why sub-2e-7 baselines correctly stay on the faithful N-ary kernel, and a seeded classification sweep. Gate mutation-tested in both directions.