Split minimizer settings from fit-result outputs

docs/dev/adrs/accepted/analysis-cif-fit-state.md

@@ -26,7 +26,7 @@ Analysis-owned fit state needs to persist:
 - pre-fit scalar snapshots for recovery workflows
 - compact status metadata for the latest saved fit projection
 - deterministic correlation summaries
-- minimizer-specific fit outputs on the active `_minimizer.*` category
+- minimizer-specific fit outputs on the paired `_fit_result.*` category
 - per-parameter posterior summaries on `_fit_parameter`
 - large posterior arrays and plot caches in `analysis/results.h5`
 
@@ -47,8 +47,7 @@ Persist analysis-owned fit state as explicit analysis categories in
 
 Do not add a dedicated `_fit_state` category or
 `_fit_state.schema_version`. Persisted fit state is detected from
-`_fit_result`, `_fit_parameter`, `_fit_parameter_correlation`, and
-fit-output fields on `_minimizer.*`.
+`_fit_result`, `_fit_parameter`, and `_fit_parameter_correlation`.
 
 ### Common fit-state categories
 
@@ -77,7 +76,8 @@ pre-fit scalar snapshots:
 - `posterior_gelman_rubin`
 - `posterior_effective_sample_size_bulk`
 
-`_fit_result` stores the latest saved fit header:
+`_fit_result` stores the latest saved fit header and scalar
+family-specific fit outputs:
 
 - `result_kind`
 - `success`
@@ -92,9 +92,9 @@ pairs are stored.
 
 ### Minimizer fit projection
 
-The active `_minimizer.*` category stores both user-selected solver
-inputs and fit-filled outputs. Deterministic minimizer classes store
-compact fit output counts:
+The active `_minimizer.*` category stores user-selected solver inputs
+only. Scalar outputs are written to the paired `_fit_result.*` category.
+Deterministic fit-result classes add compact fit output counts:
 
 - `objective_name`
 - `objective_value`
@@ -104,17 +104,14 @@ compact fit output counts:
 - `degrees_of_freedom`
 - `covariance_available`
 - `correlation_available`
-- `runtime_seconds`
-- `iterations_performed`
 - `exit_reason`
 
 Do not persist a `_deterministic_parameter_result` category. Final
 deterministic parameter values and uncertainties already persist in the
 model CIF files, and restored deterministic ordering comes from
 `_fit_parameter`.
 
-Bayesian minimizer classes store sampler inputs and fit outputs under
-`_minimizer.*`, including:
+Bayesian minimizer classes store sampler inputs under `_minimizer.*`:
 
 - `sampling_steps`
 - `burn_in_steps`
@@ -123,7 +120,9 @@ Bayesian minimizer classes store sampler inputs and fit outputs under
 - `parallel_workers`
 - `initialization_method`
 - `random_seed`
-- `runtime_seconds`
+
+Bayesian fit-result classes store scalar outputs under `_fit_result.*`:
+
 - `point_estimate_name`
 - `sampler_completed`
 - `credible_interval_inner`
@@ -170,10 +169,10 @@ posterior displays.
 Load order is:
 
 1. standard analysis configuration
-2. common fit-state categories
-3. `_minimizer.*` fit-output fields according to the active
-   `_minimizer.type`
-4. posterior sidecar arrays when a Bayesian result is expected
+2. `_minimizer.*` settings according to the active `_minimizer.type`
+3. common and family-specific `_fit_result.*` fields on the paired class
+4. `_fit_parameter` and `_fit_parameter_correlation`
+5. posterior sidecar arrays when a Bayesian result is expected
 
 Persist backend runtime objects, optimizer instances, and raw driver
 payloads nowhere in this design.

docs/dev/adrs/accepted/display-ux.md

@@ -200,6 +200,12 @@ Use these naming rules:
   path for `versus`.
 - `posterior.*` names are used only when posterior samples are required.
 
+`project.display.fit.results()` also prints a "Settings used" block
+above the result tables. The block is sourced from
+`analysis.minimizer.*` so the minimizer inputs and paired
+`analysis.fit_result.*` outputs are visible from the accepted display
+facade without adding a new `Analysis`-level display method.
+
 ## Rejected Alternatives
 
 Flat display facade:

docs/dev/adrs/accepted/minimizer-category-consolidation.md

@@ -55,31 +55,34 @@ samplers.
 
 ## Decision
 
-### 1. Unified `minimizer` category replaces all sampler-input and fit-result categories
+### 1. Unified `minimizer` category replaces sampler-input categories
 
 Introduce a single switchable category `minimizer` on `Analysis`. Its
 concrete class is determined by `Analysis.minimizer_type`. The category
-holds both user-writable inputs and fit-filled outputs in one place.
+now holds user-writable minimizer inputs only. The later
+[`minimizer-input-output-split.md`](minimizer-input-output-split.md) ADR
+reverses the fit-output half of this rule: scalar fit outputs live on
+the paired `fit_result` category instead of on `minimizer`.
 
 The following categories are removed:
 
 - `bayesian_sampler` — fields move into the Bayesian concrete classes of
   `minimizer`.
 - `bayesian_result`, `bayesian_convergence` — fields move into the
-  Bayesian concrete classes of `minimizer` (`runtime_seconds`,
+  Bayesian concrete classes of `fit_result` (`fitting_time`,
   `acceptance_rate_mean`, `gelman_rubin_max`,
   `effective_sample_size_min`, `best_log_posterior`, …).
 - `deterministic_result` — fields move into the deterministic concrete
-  classes of `minimizer` (`runtime_seconds`, `iterations_performed`,
-  `exit_reason`, …).
+  classes of `fit_result` (`fitting_time`, `iterations`,
+  `objective_value`, `exit_reason`, …).
 - `bayesian_parameter_posterior` — replaced by `Parameter.posterior`
   (see §3).
 - `bayesian_distribution_cache`, `bayesian_pair_cache`,
   `bayesian_predictive_dataset` — replaced by HDF5 sidecar (see §4).
 
-`fit_result` and `fit_parameter` (analysis-owned bounds, success flag,
-reduced chi-square, message, fit time, iterations) remain unchanged as
-fit-mode-agnostic header categories.
+`fit_parameter` (analysis-owned bounds) remains a fit-state category.
+`fit_result` remains the common fit header category and is extended by
+the input/output split ADR with family-specific scalar outputs.
 
 ### 2. Selectors move to the `Analysis` owner
 
@@ -388,10 +391,11 @@ category's class-level `_engine_metadata` dict.
 
 ### Trade-offs
 
-- `minimizer` is the first category that mixes writable user inputs and
-  writable fit-filled outputs in the same scope. This is a small new
-  convention but is the natural generalization of how `Parameter`
-  already holds both user input and refined value on the same object.
+- `minimizer` no longer mixes writable user inputs and fit-filled
+  outputs in the same scope. That stricter boundary is recorded by
+  [`minimizer-input-output-split.md`](minimizer-input-output-split.md);
+  `Parameter` remains the refinement-in-place precedent for model values
+  rather than minimizer diagnostics.
 - The set of `_minimizer.*` tags present in CIF depends on the active
   `_fitting.minimizer_type`. Loading a CIF whose tags don't match the
   minimizer's allowed set raises (clear validation, not silent
@@ -461,9 +465,10 @@ category count for each new sampler and entrenches the convention break.
 
 ### D. Strict input-only `minimizer` plus a separate `fit_result`
 
-Keep the categories single-concept (inputs xor outputs) at the cost of
-two-place lookup for related info. Rejected in favour of the
-one-category-mixes-both shape (§1, §"Trade-offs") because the existing
-`Parameter` model already mixes input and refined value on the same
-object, and one-place discoverability is more valuable than strict
-purity.
+Originally rejected in favour of the one-category-mixes-both shape (§1,
+§"Trade-offs"). Reversed by
+[`minimizer-input-output-split.md`](minimizer-input-output-split.md)
+after implementation showed the `Parameter` analogy does not hold for
+minimizer settings versus fit diagnostics. The current design keeps
+`minimizer` input-only and moves scalar fit outputs to the paired
+`fit_result` category.