Write paper: MS-TCM model and fits to FRFR-category data

## Goal

Write a paper presenting the Multi-Stream Temporal Context Model (MS-TCM) — described in [notes/ms-tcm.pdf](../blob/main/notes/ms-tcm.pdf) — together with MLE parameter estimates and 95 % confidence intervals fit to the category condition of Manning et al. (2023) FRFR.

## Arc

1. **Motivation.** The pilot equivalence in Xu, Duncan, & Manning (2026) — across-event-bridge recall statistically equivalent to across-event-within-storyline (BF₀₁ = 412) — is unexpected under standard TCM. MS-TCM extends TCM with storyline-specific context vectors that drift only when their storyline is active; this predicts the equivalence under identifiable parameter regimes.
2. **Model.** Formalize MS-TCM per `notes/ms-tcm.pdf` (global + per-storyline context updates, composite encoding, task-dependent retrieval reweighting, optional §5 mechanisms γ / α / λ). Include the §4.4 analytical anchor (composite similarity 0.866 grouped vs 0.806 bridge at β_G = β_S = 0.5, w_G = 0.2, w_S = 0.8, m = 3) as a worked example.
3. **Worked example — FRFR category condition.** Reinterpret each list's four semantic categories as four storylines (K = 4). Early lists (sorted by category) give grouped-within-storyline structure; late lists (randomly ordered) give interleaved / bridge structure within the same participants.
4. **Fits.** Maximum-likelihood estimates with 95 % bootstrap confidence intervals (resample participants with replacement, ≥1000 reps) for β_G, β_S, w_G, w_S, γ, λ, with a standard-TCM baseline (w_S = 0) for comparison. Model comparison via AIC / BIC / held-out log-likelihood.
5. **Interpretation.** Do the FRFR fits recover a storyline-context-dominant regime (w_S ≫ w_G)? How do early-list (grouped) vs late-list (interleaved) recall transitions compare under MS-TCM vs standard TCM?

## Scope

- Target venue: memory / cognitive science journal (JML, JMP, Psych Review).
- Reuses the CDL LaTeX template already in `paper/`.
- Figures reproduced from `code/notebooks/`.

## Dependencies

- Feature 001 (model + FRFR-category dataset + MLE-with-CI fitter) must be complete.
- Then: figure notebooks, prose, references (CDL-bibliography submodule).

## Priority

Scientific endpoint of the project. Blocked on feature 001.