research: add SYM-ARCH-002A6 online measurement dynamics

[Tristan-Stoltz-ERC]

Summary

Adds SYM-ARCH-002A6 online measurement dynamics as a draft sibling tranche directly on the frozen 002A core (#57).

This PR adds no architecture mechanism and consumes no CONFIRM data. It closes measurement gaps exposed by the negative SYM-ARCH-001 result, especially the ambiguity between low forgetting and weak acquisition.

Prequential acquisition contract

Every online-learning step is measured in this order:

1. predict before consuming the current label;
2. record correctness and inference latency;
3. apply the labeled update;
4. record update latency;
5. record post-update resource state.

Post-update correctness on the same item is not admissible as acquisition evidence.

Provenance-bound phase traces

Each OnlineMeasurementTrace binds:

• experiment-manifest digest;
• task-program digest;
• runtime-context digest;
• contiguous phase identity;
• optional initial learner/spec-state digest;
• ordered prequential steps.

The trace receives its own domain-separated BLAKE3 digest.

Acquisition metrics

A6 adds:

• sustained examples-to-criterion using a frozen rolling window, accuracy threshold, and required consecutive qualifying windows;
• overall prequential accuracy;
• terminal-window accuracy;
• normalized cumulative-accuracy AUC that rewards earlier acquisition at matched overall accuracy.

The cumulative-AUC definition is frozen and cannot be silently replaced by ordinary mean correctness or post-update accuracy.

Criterion validity

experiment_measurement_validity audits the frozen acquisition criterion against a preregistered simple-reference accuracy ceiling and a minimum practical excess.

It exposes:

• integer correct count actually required by the finite window;
• effective finite-window accuracy threshold;
• effective excess above the reference ceiling;
• admissible vs reference_confounded status with qualifiers.

A claim-bearing examples-to-criterion result may not use a reference_confounded criterion.

This is a construct-validity guard only; A4 shortcut controls and A2 uncertainty/power remain required.

Latency / throughput

For inference and update latency independently:

• sample count;
• total nanoseconds;
• mean;
• p50;
• p95;
• observations/second when measured time is nonzero.

Latency traces are runtime-context-bound. A6 does not authorize cross-hardware latency claims from unmatched runtime-context digests.

Resource trace

Post-update snapshots record:

• trainable scalar parameters;
• persistent state bytes;
• replay bytes;
• temporal state bytes;
• optional RSS.

Summaries report final/peak trainable and persistent state plus peak replay, temporal state, and observed RSS.

Scientific boundary

A6 complements rather than replaces the continual R[t_train][t_eval] matrix:

• R matrix → retention/transfer across tasks;
• A6 → acquisition/recovery dynamics inside each phase.

A system with low forgetting but slow/absent criterion attainment can therefore no longer be described as strong continual learning merely because retention-style metrics look stable.

CI

Dedicated exact-head workflow gates:

• cargo fmt --all -- --check;
• focused experiment_measurement tests;
• focused experiment_measurement_validity tests;
• full symthaea-psych-bench library check on Rust 1.96.

Wording ceiling

Merging this PR supports only:

Symthaea psych-bench contains provenance-bound prequential acquisition, criterion-validity, latency/throughput, and resource measurement primitives suitable for later preregistered continual-learning experiments.

It does not support a Symthaea performance claim, a faster-learning claim, a hardware-independent latency claim, a resource-efficiency claim, or a claim-bearing threshold that fails the reference audit.

Tracks #55. Depends only on #57.