fix(harness): I() wrap t_observed/observed_dv too; symmetric scalar coercion

The previous sims_at_observed I() wrap (commit 41a2f95) closed only
one third of the bug class. The same auto_unbox-without-I() collapse
hits ``t_observed`` and ``observed_dv`` on n_obs == 1 subjects:
``jsonlite::toJSON(..., auto_unbox = TRUE)`` of a length-1 numeric
vector yields a JSON scalar (``57.5``) rather than a length-1 array
(``[57.5]``), and ``PredictedSimulationsSubject.t_observed:
list[float]`` rejects the float with ``Input should be a valid list``.

Surfaces on every n_obs == 1 subject in pheno_sd; aborts the entire
fixture before any per-fold scorecard is produced.

Two-layer fix (symmetric with the existing sims_at_observed wire)
-----------------------------------------------------------------
- **R-side primary** (r/harness.R): wrap ``t_observed = I(t_obs)``
  and ``observed_dv = I(dv_obs)`` in ``.simulate_posterior_predictive``
  so jsonlite preserves array shape under auto_unbox = TRUE.
- **Pydantic-side defence in depth** (backends/r_schemas.py): a new
  ``field_validator("t_observed", "observed_dv", mode="before")``
  coerces a bare scalar to ``[float(v)]`` so a future R-side
  regression fails the harness pin (see below) instead of the
  Phase-1 weekly run.

Tests
-----
2 new pins in ``TestPredictedSimulations1DCoercion``:
- ``test_harness_uses_I_wrap_for_sparse_subjects`` extended to also
  pin ``t_observed = I(t_obs)`` and ``observed_dv = I(dv_obs)``
  substrings in r/harness.R.
- ``test_scalar_t_observed_is_coerced_to_list`` exercises the full
  n_obs == 1 auto_unbox-without-I() shape (bare floats for both
  obs vectors + flat list for sims) and asserts all three coerce.

Bundled changes (unrelated, were already in the working tree)
-------------------------------------------------------------
- README, CHANGELOG, docs (ARCHITECTURE, FORMULAR, suite_a/c READMEs,
  src/r README): documentation drift cleanup that landed alongside.
- src/apmode/cli.py, bundle/models.py, dsl/stan_emitter.py: ruff /
  formatting touch-ups.
- tests/unit/test_deep_inspect.py: matching test updates.

Lint + tests
------------
``ruff check`` + ``ruff format`` clean across src/ + tests/.
``mypy --strict`` clean over 136 source files. Full non-live sweep:
2551 passed, 4 skipped, 17 deselected, 15 warnings in 135 s.

Author: biostochastics

CHANGELOG.md

@@ -7,6 +7,48 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+### Fixed — Documentation accuracy and drift cleanup
+
+- Synced README / CLAUDE auto-managed counters to the current test collection
+  and clarified the CLI-command count versus registered command groups.
+- Reconciled HTTP API security documentation with the implemented static
+  API-key and dataset-root requirements for non-loopback binds.
+- Updated Architecture, Formular, Suite A/C, and R adapter docs for the
+  v0.6.1-rc1 surface: 10 Formular transforms, shipped API/RO-Crate commands,
+  v0.7 absorption preview forms, BLQ M3/M4 Stan support, and the current
+  benchmark fixture rosters.
+- Removed stale/broken ADR references and repo-resident process attribution
+  from documentation and changelog prose.
+
+### Fixed — CLI trace/graph hardening: typed traces, valid exporters, DAG invariants
+
+The `apmode trace` and `apmode graph` deep-inspection commands now handle
+malformed or adversarial bundle artifacts without corrupting machine-readable
+output or crashing read-only inspection:
+
+- `apmode trace --json` keeps stdout as pure JSON even when
+  `agentic_iterations.jsonl` contains corrupt lines. Invalid iteration rows are
+  validated against `AgenticIterationEntry` and skipped with human-mode
+  diagnostics rather than causing renderer crashes.
+- `apmode trace --cost` now tolerates malformed token/cost/time fields in
+  `iter_*_meta.json`, warning in human mode and using zero for bad numeric
+  cells instead of raising `ValueError`.
+- `apmode graph --format dot` and `--format mermaid` now emit raw text on
+  stdout. Rich no longer interprets DOT attributes or Mermaid labels as markup
+  (for example `node [shape=box, fontsize=10]` is preserved).
+- Mermaid export uses stable generated node IDs (`n0`, `n1`, ...) so candidate
+  IDs such as `a-b` and `a_b` cannot collide after sanitization; original
+  candidate IDs remain visible in labels.
+- `SearchGraph` validation now rejects duplicate `candidate_id` values and
+  edges that reference missing nodes, aligning the bundle model with the CLI's
+  advertised DAG semantics.
+- `apmode graph --backend` help now includes `bayesian_stan`, matching the
+  accepted backend enum.
+
+Tests: `tests/unit/test_deep_inspect.py` adds regression pins for corrupt trace
+rows under `--json`, invalid typed iteration rows, malformed cost metadata, DOT
+stdout, Mermaid ID collisions, duplicate graph nodes, and dangling graph edges.
+
 ### Fixed — Sparse-data harness: 1D `sims_at_observed` no longer crashes the runner
 
 When a subject has a single observation (the `pheno_sd` neonatal
@@ -152,9 +194,9 @@ SumIG closed-form input rate integrates to ≈1 over (0, ∞).
 
 ### Fixed — Multi-CLI code review hardening pass: cancellation, durability, digest scoping, harness diagnostics
 
-A multi-model code review (droid, crush=glm-5, gemini-3-pro, opencode=minimax-m2.7) of the
-`api/`, `bundle/`, `backends/`, and `bayes/` surfaces produced 9 actionable findings, all
-addressed in this set of patches. No public API change; existing bundles remain loadable.
+A focused code review of the `api/`, `bundle/`, `backends/`, and `bayes/`
+surfaces produced 9 actionable findings, all addressed in this set of patches.
+No public API change; existing bundles remain loadable.
 
 **API run-state cancellation contract — slot leak on double-cancel closed.**
 `apmode.api.runs.execute_run` now wraps the `on_complete` callback and the `FAILED`
@@ -562,11 +604,10 @@ and SAEM streaming progress.
   `--allow-public` test was updated to satisfy the concurrent
   `APMODE_API_KEY` + `--dataset-root` security gates.
 
-### Fixed — Multi-model audit pass: cancellation, auth, digest integrity
+### Fixed — Audit pass: cancellation, auth, digest integrity
 
-A targeted hardening sweep informed by an external multi-model review
-(self + clink: gemini, droid, crush, opencode). All 2035 non-live unit
-tests pass; `mypy --strict` and `ruff check` are clean.
+A targeted hardening sweep of cancellation, auth, and digest-integrity surfaces.
+All 2035 non-live unit tests pass; `mypy --strict` and `ruff check` are clean.
 
 - **Bayesian Gate 1 no longer clobbers the classical Gate 1 audit
   artifact.** `BundleEmitter.write_gate_decision` now accepts

README.md

@@ -18,6 +18,10 @@ Requires: R 4.4+, rxode2, jsonlite, lotri
 | a2_2cmt_iv_parallel_mm.csv | 2-cmt IV, parallel linear+MM elim | Compartment count + nonlinear CL |
 | a3_transit_1cmt_linear.csv | Transit (n=3), 1-cmt, linear elim | Transit chain detection |
 | a4_1cmt_oral_mm.csv | 1-cmt oral, MM elimination | Nonlinear clearance detection |
+| a5_tmdd_qss.csv | TMDD quasi-steady-state (SC mAb) | TMDD vs. 2-cmt confusion |
+| a6_1cmt_covariates.csv | 1-cmt oral + allometric WT + renal covariate | Covariate structure recovery |
+| a7_2cmt_node_absorption.csv | 2-cmt + NODE saturable absorption | NODE shape recovery + surrogate fidelity |
+| a8_1cmt_tvcl_covariate.csv | 1-cmt oral + time-varying CL + CRCL covariate | Time-varying clearance + covariate recovery |
 
 ## Reference Parameters
 

benchmarks/suite_a/README.md

@@ -3,10 +3,10 @@
 > APMODE plan Tasks 38–43, blueprint in `docs/plans/2026-04-24-apmode-v0.6-completion.md`.
 
 Phase-1 of Suite C compares APMODE's PK fits against published, peer-reviewed
-reference parameterizations on real clinical datasets. Each fixture pairs a
+reference parameterizations and selected ground-truth reference datasets. Each fixture pairs a
 NONMEM-style CSV (resolved via `dataset_id` against
 `benchmarks/datasets/registry.yaml`) with a `DSLSpec` JSON and a
-`LiteratureFixture` YAML capturing the published parameter values plus the
+`LiteratureFixture` YAML capturing the reference parameter values plus the
 parameterization mapping needed to translate published symbol names
 (e.g. `TVCL`) into APMODE's DSL-canonical names (e.g. `CL`).
 
@@ -16,8 +16,10 @@ parameterization mapping needed to translate published symbol names
 |------------|-------|--------------|---------------|
 | `theophylline_boeckmann_1992` | oral | 1-cmt + FO ka | [10.1002/psp4.12471](https://doi.org/10.1002/psp4.12471) (Schoemaker et al. 2019, nlmixr SAEM/FOCEI grid which contains the Theoph fit) |
 | `warfarin_funaki_2018` | oral | 1-cmt + lagged-FO ka | [10.1002/psp4.12445](https://doi.org/10.1002/psp4.12445) (Fidler et al. 2019, nlmixr documentation including Holford warfarin) |
-| `mavoglurant_wendling_2015` | oral | 2-cmt + FO ka | [10.1007/s11095-014-1574-1](https://doi.org/10.1007/s11095-014-1574-1) (Wendling et al. 2015, mavoglurant population PK; APMODE uses the simpler 2-cmt approximation since the published sum-of-inverse-Gaussian input function is outside DSL scope) |
-| `gentamicin_germovsek_2017` | iv_bolus | 1-cmt | [10.1128/AAC.00481-17](https://doi.org/10.1128/AAC.00481-17) (Germovsek et al. 2017, gentamicin IOV neonates) |
+| `mavoglurant_wendling_2015` | oral | 2-cmt + FO ka | [10.1007/s11095-014-1574-1](https://doi.org/10.1007/s11095-014-1574-1) (Wendling et al. 2015, mavoglurant population PK; the Phase-1 fixture uses the simpler 2-cmt approximation, while SumIG absorption remains a v0.7 preview path) |
+| `phenobarbital_grasela_1985` | iv_bolus | 1-cmt | [10.1159/000457062](https://doi.org/10.1159/000457062) (Grasela & Donn 1985, phenobarbital neonatal PK) |
+| `oral_1cpt_acop_2016` | oral | 1-cmt + FO ka | [10.32614/CRAN.package.nlmixr2data](https://doi.org/10.32614/CRAN.package.nlmixr2data) (ACOP 2016 / `nlmixr2data::Oral_1CPT` ground-truth recovery fixture) |
+| `gentamicin_germovsek_2017` | iv_bolus | 1-cmt | [10.1128/AAC.02659-16](https://doi.org/10.1128/AAC.02659-16) (Germovsek et al. 2017, gentamicin IOV neonates) |
 | `schoemaker_nlmixr2_tutorial` | iv_bolus | 1-cmt | [10.1002/psp4.12471](https://doi.org/10.1002/psp4.12471) (Schoemaker et al. 2019, `pkBolus1cmt` reference values) |
 
 ## Why the names changed from the original plan draft

benchmarks/suite_c/README.md

@@ -1,8 +1,8 @@
 # APMODE Technical Architecture
 
 **Version:** 0.3
-**Date:** 2026-04-17
-**Status:** Current (tracks APMODE 0.5.0-rc2; Phase 3 in progress)
+**Date:** 2026-04-25
+**Status:** Current (tracks APMODE 0.6.1-rc1; Phase 3 in progress)
 **Derived from:** PRD v0.3 (§3–§8)
 **Supersedes:** v0.2 (2026-04-13). Change summary: Phase 2/3 framing removed where
 shipped (Bayesian backend, NODE, agentic LLM, Gate 2.5, FREM, Gate 3 ranking all
@@ -19,8 +19,9 @@ Flowcept) moved to §10 Non-goals.
    + subprocess) is the main loop. The `BackendRunner` protocol is designed so Flyte
    or Temporal could be swapped in for GPU scheduling without rewriting adapters, but
    no such migration is planned (see §10).
-2. **Process isolation is non-negotiable.** Every backend (R/nlmixr2, Python/JAX,
-   LLM SDK, R/Rstan) runs in its own subprocess. A segfault in R must not crash the
+2. **Process isolation is non-negotiable.** Every backend boundary (R/nlmixr2,
+   Python/JAX, LLM SDK calls, cmdstanpy/CmdStan) is isolated from the orchestrator
+   where practical. A segfault in R must not crash the
    orchestrator.
 3. **The DSL is the moat.** All model specifications flow through the typed PK DSL
    (Formular). No backend emits or consumes model code outside of a deterministic
@@ -79,7 +80,7 @@ See `docs/FORMULAR.md` for the language reference.
 | AST canonicalization | `src/apmode/dsl/normalize.py` |
 | Semantic validator | `src/apmode/dsl/validator.py::validate_dsl(spec, lane=...)` — lane-aware, non-fail-fast |
 | Priors + admissibility | `src/apmode/dsl/priors.py` (families, target taxonomy, `_VALID_FAMILIES` matrix) |
-| Transforms | `src/apmode/dsl/transforms.py` (6) + `prior_transforms.py::SetPrior` (1) — 7 total |
+| Transforms | `src/apmode/dsl/transforms.py` (9) + `prior_transforms.py::SetPrior` (1) — 10 total |
 | Emitters | `nlmixr2_emitter.py`, `stan_emitter.py`, `frem_emitter.py` |
 | Testing | Hypothesis property tests + syrupy golden masters for emitter output |
 
@@ -93,7 +94,7 @@ See `docs/FORMULAR.md` for the language reference.
 | R invocation | `Rscript src/apmode/r/harness.R` via subprocess; file-based I/O (§4.2) |
 | Stan invocation | `cmdstanpy.CmdStanModel` via `src/apmode/bayes/harness.py` wrapped by `bayesian_runner.py` |
 | Retry/timeout | Bespoke logic in each runner; timeout from policy file, killed attempts write to new `attempt_id/` subdir |
-| CLI | Typer (`src/apmode/cli.py`) — 15 commands |
+| CLI | Typer (`src/apmode/cli.py`) — 16 direct commands plus registered `bundle` / `completion` groups |
 
 **Note on deployment posture.** The codebase runs natively on the user's machine via
 `uv`; there is no Docker-Compose stack, no R container, no K8s. Users who want
@@ -212,8 +213,8 @@ here.
 
 GitHub Actions: `uv sync --all-extras` → pytest → mypy strict → ruff check + format.
 Pre-commit runs ruff + mypy + the policy validator. Matrix is Python 3.12 / 3.13 /
-3.14. The benchmark suites run as separate workflows (Suite A nightly, B weekly, C
-quarterly — cadence tracked in `.github/workflows/`).
+3.14. Suite C has a dedicated workflow in `.github/workflows/suite_c_phase1.yml`;
+other benchmark cadences are operator-driven unless a workflow exists in `.github/workflows/`.
 
 ---
 
@@ -222,7 +223,7 @@ quarterly — cadence tracked in `.github/workflows/`).
 ```
               ┌────────────────────────────────────────────┐
               │            apmode CLI (Typer)              │
-              │    15 commands: run | validate | inspect    │
+              │    16 direct commands + registered groups    │
               │    datasets | explore | diff | log | trace  │
               │    lineage | graph | report | doctor | ls   │
               │    policies | bundle                        │
@@ -293,11 +294,11 @@ All paths rooted at `src/apmode/`.
 | `dsl/ast_models.py` | `DSLSpec` + all module Pydantic nodes |
 | `dsl/normalize.py` | AST canonicalization |
 | `dsl/validator.py` | Lane-aware `validate_dsl` |
-| `dsl/transforms.py` | 6 structural transforms + `FormularTransform` union |
-| `dsl/prior_transforms.py` | `SetPrior` transform (7th) |
+| `dsl/transforms.py` | 9 structural transforms + `FormularTransform` union |
+| `dsl/prior_transforms.py` | `SetPrior` transform (10th) |
 | `dsl/priors.py` | Prior families + `_VALID_FAMILIES` admissibility matrix |
 | `dsl/nlmixr2_emitter.py` | AST → nlmixr2 R code |
-| `dsl/stan_emitter.py` | AST → Stan program (BLQ M3/M4, IOV, NODE ⇒ `NotImplementedError`) |
+| `dsl/stan_emitter.py` | AST → Stan program (IOV, NODE, maturation covariates, and v0.7 absorption preview forms ⇒ `NotImplementedError`) |
 | `dsl/frem_emitter.py` | AST → FREM-augmented nlmixr2 |
 | `dsl/_emitter_utils.py` | Shared emitter helpers |
 
@@ -369,7 +370,7 @@ All paths rooted at `src/apmode/`.
 
 | File | Role |
 |------|------|
-| `cli.py` | Typer app — 15 commands |
+| `cli.py` | Typer app — 16 direct commands plus `bundle` / `completion` groups |
 | `paths.py` | `APMODE_POLICIES_DIR` env override + pyproject-walk fallback |
 | `routing.py` | Lane Router — evidence-manifest-driven dispatch |
 | `logging.py` | `structlog` configuration |
@@ -613,9 +614,9 @@ Phases 1 and 2 are complete. Phase 3 is in progress per CLAUDE.md. The per-month
 task list from v0.2 has been removed from this doc; it is preserved in the git
 history at `docs/ARCHITECTURE.md@v0.2` and summarized in `CHANGELOG.md`.
 
-**What is active today (0.5.0-rc2):**
+**What is active today (0.6.1-rc1):**
 
-- DSL grammar + compiler + validator + 7 typed transforms.
+- DSL grammar + compiler + validator + 10 typed transforms.
 - Classical NLME backend (nlmixr2, SAEM/FOCEi) with warm-start children.
 - Bayesian backend (Stan / Torsten via `cmdstanpy`) with R̂ / ESS / E-BFMI /
   Pareto-k Gate 1 integration.
@@ -625,13 +626,15 @@ history at `docs/ARCHITECTURE.md@v0.2` and summarized in `CHANGELOG.md`.
 - FREM + MI-PMM + MI-missRanger missing-data pipelines.
 - Gate 1 (PIT / NPDE-lite) + Gate 2 (lane-specific) + Gate 2.5 (ICH M15) +
   Gate 3 (Borda / weighted ranking with uniform-drop rule).
-- Reproducibility bundle with `_COMPLETE` sentinel.
-- 14-command CLI + HTML / Markdown regulatory report.
-- Benchmark Suite A (full, 7 scenarios) + B + C scaffolding.
+- Reproducibility bundle with `_COMPLETE` sentinel and RO-Crate projection.
+- Typer CLI (`run`, bundle inspection/reporting, HTTP `serve`, RO-Crate/SBOM subcommands) + HTML / Markdown regulatory report.
+- Benchmark Suite A (8 scenarios), Suite B perturbation anchors, and Suite C literature fixtures.
+- FastAPI HTTP surface (`apmode serve`) with loopback default, static API-key floor for non-loopback binds, SQLite run store, and cancellation lifecycle.
 
-**What remains for Phase 3 completion:** LORO-CV full integration into the
-Optimization lane, NODE posterior-predictive simulation (currently inert stub),
-Stan-side BLQ M3/M4 + IOV + maturation-covariate lowering, REST API.
+**What remains for Phase 3 completion:** NODE posterior-predictive simulation
+(currently inert stub), Stan-side IOV + maturation-covariate lowering, full
+Stan/Torsten support for the v0.7 absorption preview forms, and broader
+production hardening around public API deployments.
 
 ---
 
@@ -671,9 +674,7 @@ initial-estimate strategy: NCA-seeded + warm-start per §2.5). Items still open:
    and a pharmacometric-review workflow. PRD §10 Q1.
 2. **Regulatory engagement timing** — when to seek informal FDA / EMA feedback on
    the credibility framework (PRD §10 Q4).
-3. **LORO-CV operationalization** — the Gate 2 policy fields are in place but the
-   Optimization-lane pipeline wiring is in progress (Phase 3 completion).
-4. **NODE posterior-predictive simulation** — `NodeBackendRunner.sample_posterior_predictive`
+3. **NODE posterior-predictive simulation** — `NodeBackendRunner.sample_posterior_predictive`
    is a discoverable inert stub; Phase 3 completion item. Concrete integration
    point: `backends/node_trainer.py`.
 
@@ -693,7 +694,6 @@ this section is the canonical record of why they were deferred.
   `structlog` JSON output.
 - **Rust parser migration.** Lark in Python is adequate; the Phase-2 LALRPOP+PyO3
   branch was considered and rejected.
-- **REST API.** Planned for Phase 3 completion — out of scope for 0.5-rc.
 - **Web UI.** A minimal browser UI was prototyped and removed; the CLI + HTML
   report is the current UX.
 

docs/ARCHITECTURE.md

@@ -1,7 +1,7 @@
 # Formular — The APMODE PK Specification Language
 
-**Version:** 0.5 (tracks APMODE 0.5.0-rc1)
-**Status:** Current (Phase 3 in progress)
+**Version:** 0.6.1 (tracks APMODE 0.6.1-rc1)
+**Status:** Current
 **Canonical source:** `src/apmode/dsl/` — grammar (`pk_grammar.lark`), AST (`ast_models.py`),
 validator (`validator.py`), transforms (`transforms.py` + `prior_transforms.py`), priors
 (`priors.py`), emitters (`nlmixr2_emitter.py`, `stan_emitter.py`, `frem_emitter.py`).
@@ -266,12 +266,14 @@ Enforced at `SetPrior` validation time (`apmode.dsl.priors::_VALID_FAMILIES`). I
 ### Provenance and FDA alignment
 
 Every `PriorSpec` carries a `source ∈ {uninformative, weakly_informative,
-historical_data, expert_elicitation, meta_analysis}`. Sources other than
-`uninformative` / `weakly_informative` require a non-empty `justification` string
-(validator-enforced); `historical_data` additionally requires `historical_refs` listing
-source datasets. This aligns with FDA draft guidance FDA-2025-D-3217 (January 2026) on
-Bayesian methodology — the resulting `prior_manifest.json` in the bundle is the
-artifact Gate 2 consumes for prior-justification review.
+historical_data, expert_elicitation, meta_analysis, fixed_external}`. The
+informative sources (`historical_data`, `expert_elicitation`, `meta_analysis`) require
+a non-empty `justification` string (validator-enforced); `historical_data` additionally
+requires `historical_refs` listing source datasets. `fixed_external` marks externally
+fixed parameter values used by the SumIG disposition-fixed gate. This aligns with FDA
+draft guidance FDA-2025-D-3217 (January 2026) on Bayesian methodology — the resulting
+`prior_manifest.json` in the bundle is the artifact Gate 2 consumes for
+prior-justification review.
 
 ---
 
@@ -515,9 +517,10 @@ def rhs(t, y, theta):
 
 The emitters are in `nlmixr2_emitter.py`, `stan_emitter.py`, and the NODE harness in
 `backends/node_*.py`. Any spec that passes `validate_dsl` for the target lane will
-lower cleanly in nlmixr2; Stan additionally requires that BLQ M3/M4 and IOV are
-absent (Phase 3 work) and that NODE modules are not present (Stan has no neural-ODE
-support).
+lower cleanly in nlmixr2. Stan supports the core mechanistic forms and BLQ M3/M4
+censored likelihoods; it still rejects IOV, maturation covariate links, NODE modules,
+and the v0.7 absorption preview forms (`Erlang`, `ParallelFirstOrder`, `SumIG`) pending
+follow-on lowering work.
 
 ---
 

docs/FORMULAR.md

@@ -101,6 +101,25 @@ class PredictedSimulationsSubject(BaseModel):
     # reshapes into a numpy (n_sims, n_obs) array before validation.
     sims_at_observed: list[list[float]] = Field(min_length=1)
 
+    @field_validator("t_observed", "observed_dv", mode="before")
+    @classmethod
+    def _coerce_scalar_to_list(cls, v: object) -> object:
+        """Tolerate a bare scalar for the n_obs == 1 case.
+
+        The R harness wraps single-observation vectors with ``I(...)``
+        so ``jsonlite::toJSON(..., auto_unbox = TRUE)`` keeps length-1
+        numeric arrays as JSON arrays. Without the wrap (older bundles
+        or a future regression), a subject with n_obs == 1 emits
+        ``t_observed: 1.5`` instead of ``t_observed: [1.5]`` and the
+        ``list[float]`` field rejects the float with
+        ``Input should be a valid list``. Coerce a bare scalar to a
+        length-1 list defensively; well-formed inputs (list-of-floats)
+        are unaffected.
+        """
+        if isinstance(v, (int, float)) and not isinstance(v, bool):
+            return [float(v)]
+        return v
+
     @field_validator("sims_at_observed", mode="before")
     @classmethod
     def _coerce_1d_to_2d(cls, v: object) -> object: