Add joint MAP hierarchical CT-AR Rasch slice (stacked on #976)

[seonghobae]

Dependency

This PR is stacked on #976 (cursor/rust-longitudinal-state-4683). It must not be merged to main until #976 lands. It does not relabel the #976 OLS / caller-supplied AR kernels as random effects.

Duplicate check

Inspected open PRs/issues before starting, including #978, #979, and #981. No open PR implements jointly estimated hierarchical longitudinal IRT. #976 is the OLS/AR base; closed #848 is the same OLS/AR tree.

What this adds

Smallest scientifically valid jointly estimated longitudinal latent-state IRT slice:

logit P(Y_pti = 1) = theta_pt - b_i,   sum_i b_i = 0
theta_p,1 ~ N(mu, tau^2)
theta_p,t | theta_p,t-1 ~ N(
    mu + exp(-lambda * Delta_pt) * (theta_p,t-1 - mu),
    tau^2 * (1 - exp(-2 * lambda * Delta_pt))
)

Delta_pt is elapsed days from exact millisecond offsets.

Honest labels:

• estimand_scope = joint_map_hierarchical_ctar_rasch
• transition_kind = continuous_time_ar1_ou
• interval_kind = wald_measurement_observed_information
• population_random_effects_estimated = True (shared mu, tau; states shrunk toward mu)
• ar_coefficient_source = joint_map
• multiple_membership_estimated = False
• gpu_parity = False

This is joint MAP, not Fox & Glas Gibbs, not Jeon & Rabe-Hesketh adaptive quadrature, not independent OLS, and not caller-supplied discrete AR.

Exclusions (contracted)

• Multiple-membership / crossed u_h: not in this joint likelihood. Next step is a dedicated identification study combining Fox & Glas multilevel IRT with Browne et al. MMMC.
• GPU: the existing wgpu path owns MLSIRM distance/likelihood kernels, a different estimand. No GPU kernel is claimed.

Recovery honesty

Multi-seed tests recover known person-occasion states (RMSE / measurement-information Wald coverage) and the population mean. Joint MAP shrinks tau. Short irregular series leave lambda weakly identified; tests require a finite positive decay and unit-day phi in (0, 1), plus irregular-gap phi ordering, not tight unbiased-ML recovery of lambda.

Docs

ADR-0019 (Proposed), doctoring, ARCHITECTURE, traceability, PRD-FR-061 / TRD-MLT-004–005, VV-SCI-008, CHANGELOG. APA 7th primary citations; journal PDFs are copyrighted so they are cited + linked + summarized, not attached.

Evidence on this head

• cargo test --locked --manifest-path crates/mlsirm-core/Cargo.toml --lib longitudinal_irt: 11 passed
• cargo test --locked --workspace: passed
• cargo test --locked --manifest-path crates/fast-mlsirm-py/Cargo.toml: 9 passed
• pytest hierarchical IRT + longitudinal OLS/AR + loader + architecture/changelog/docs: 17 + supporting tests passed
• Rust default backend assertion: resolve_backend(auto) == rust
• Existing Atheris fuzz_config harness: 8s bounded run completed without crash
• Existing GPU-smoke test test_marginal_gpu_agrees_with_cpu_loosely: skipped here (no Vulkan adapter). This slice adds no GPU kernel; CI gpu-smoke remains the MLSIRM path and must not be claimed from this environment.

  

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro Plus

Run ID: d01be734-7247-43cf-ae8f-0c94b6e8e168

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[cursor]

Bugbot is not enabled for your account, so this pull request was not reviewed.

Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs.