research: add SYM-ARCH-002B1 strong-simple analytic baselines

[Tristan-Stoltz-ERC]

Summary

Adds SYM-ARCH-002B1 strong-simple analytic baselines for the architecture-discrimination program in #55.

This is a draft sibling of #60/#61, stacked directly on #59 / A3. It does not consume CONFIRM data and carries no architecture-performance claim.

The goal is to attack a simple alternative explanation before adding GRU, trainable SSM, Mamba, liquid dynamics, or Hebbian plasticity:

Is an apparent architecture advantage already explained by a fixed representation plus a strong online analytic readout?

Baseline ladder

B1 adds three deterministic conditions:

1. one_hot_rls — normalized categorical one-hot representation + full-covariance online RLS.
2. fixed_random_tanh_rls — deterministic fixed random projection + tanh/L2 normalization + the same RLS algorithm/configuration.
3. vanilla_hdc_rls — deterministic role/value HDC binding/bundling + normalization + the exact same RLS readout shape as the random condition at matched encoded dimension.

MatchedBaselineFamilySpec emits all three from one schema/RLS contract.

Important terminology hardening: sharing the same RLS algorithm/configuration does not automatically mean two conditions are readout-shape matched. One-hot usually has a smaller effective feature dimension, so its trainable RLS state and O(d^2) covariance can differ from random/HDC.

Executable contrast-fairness audit

experiment_baseline_fairness records exact spec digests, learner-visible schema equality, RLS contract equality, effective feature/readout/covariance shape, and paired representation-seed index when both encoders are randomized.

It returns only:

• representation_level when a representation-only interpretation is admissible;
• reference_only when capacity/protocol differs enough that such attribution is not clean.

Random↔HDC is the intended representation-level B1 contrast. Dimension-mismatched one-hot↔random/HDC is explicitly a lower-complexity reference/capacity comparison. Therefore random > one_hot is not, by itself, proof that nonlinear random expansion caused the improvement if RLS capacity also increased.

The audit does not require equal fixed-encoder storage; resource efficiency is reported separately.

Matched data and update opportunities

experiment_baseline_panel::MatchedBaselinePanel is the preferred claim-bearing execution path.

It owns all three baseline agents and:

• applies each labeled training item to all three atomically using clone-before-commit semantics;
• records an order-sensitive training-stream digest;
• evaluates all three on the same labeled evaluation item without model-state mutation;
• records a separate order-sensitive evaluation-stream digest including the expected label;
• validates that every member update count equals the panel training-observation count;
• captures member spec digests and resource footprints;
• emits a canonical panel-snapshot digest independent of member serialization order.

If any model rejects a training or evaluation item, no exposure ledger is partially committed. Claim-bearing B1 results should bind the valid panel snapshot digest; a hand-written loop over independent agents is not sufficient evidence for a matched-stream claim.

RLS / resource contract

OnlineRlsBinary uses {-1,+1} targets and standard recursive least squares with frozen ridge/forgetting/bias policy, no replay buffer, no learned encoder, and no temporal state.

Replay-free is not memory-free: full f64 inverse covariance is O(d^2) state. BaselineResourceFootprint reports feature dimension, encoder bytes, readout weights, covariance bytes, total persistent state, trainable parameters, replay examples (0), and temporal-state bytes (0).

A hard 512 MiB covariance safety ceiling rejects oversized full-RLS requests before allocation. Full RLS therefore cannot be silently applied at Symthaea's ordinary 16K HDC dimension.

Frozen schema / provenance

The categorical schema is strict: normalized/sorted unique names, explicit sorted domains, exact learner-visible feature set, fail-closed out-of-domain values, and normalized one-hot representation.

Each emitted baseline spec has a domain-separated BLAKE3 digest over its versioned configuration. That digest does not pretend to bind a git commit; exact source identity is separately frozen in the experiment manifest code_revision and later claim/evidence binding.

Scientific interpretation boundary

• HDC ≈ random under a passing fairness audit → no evidence beyond generic fixed random features.
• HDC > random under a passing fairness audit → representation-level evidence for the factorized HDC implementation.
• random > HDC → HDC is less useful in that regime.
• one-hot comparisons remain simpler-model/reference contrasts whenever the fairness audit finds a capacity mismatch.

An HDC win does not isolate HDC binding algebra. Exact continuous-vs-bipolar/BinaryHV binding attribution remains a later matched factorization/algebra tranche.

The implementation is RanDumb/F-OAL-inspired, not an exact reproduction of either external method.

Tests / CI

The dedicated workflow now gates:

• rustfmt for all B1 Rust paths;
• baseline mechanics/RLS/resource tests;
• contrast-fairness tests;
• matched-panel stream determinism, atomicity, update-count, label-binding, and canonical snapshot-digest tests;
• full psych-bench library check.

Wording ceiling

Merging this PR supports only:

Symthaea psych-bench contains deterministic, resource-audited one-hot, fixed-random, and vanilla-HDC online-RLS baselines plus executable contrast-fairness and matched-exposure provenance suitable for later preregistered comparisons.

It does not support a performance result, an HDC-vs-random conclusion, a representation-only interpretation for a reference_only contrast, a matched-stream claim without a valid panel snapshot, a named-method reproduction claim, or any conclusion about liquid/Hebbian mechanisms.

Depends on #59. Sibling of #60 and #61. Tracks #55.

[Tristan-Stoltz-ERC]

Static scientific-fairness audit hardening completed on the current B1 branch.

Two distinct gaps were closed:

1. Contrast interpretation: sharing one RLS implementation does not make one-hot↔random/HDC a readout-shape-matched contrast when effective feature dimensions differ. experiment_baseline_fairness now audits schema, RLS protocol, effective readout/covariance shape, and paired randomized seed index, returning only representation_level or reference_only. Random↔HDC is the intended clean representation contrast in the matched family; dimension-mismatched one-hot comparisons are explicitly capacity/reference comparisons.

2. Matched exposure: static specs do not prove that a runner actually showed each baseline the same data/order/update opportunities. MatchedBaselinePanel now feeds training observations to all three agents atomically, records separate order-sensitive training/evaluation stream digests, validates equal update counts, records member spec/resource snapshots, and exposes a canonical panel-snapshot digest for evidence binding. Failed training/evaluation items leave the panel ledger unchanged.

The B1 workflow now runs rustfmt plus focused tests for baseline mechanics, fairness auditing, and the matched panel before the psych-bench library check.

No performance result or HDC advantage is claimed by these changes.