[MarkusNeusinger]

Summary

• core/adaptability.py, the stdlib-only checker of the adaptable-code convention, implements rules 1 to 6 with the amendments of 2026-10-23. It returns a record {version, status, findings, block, constants, color_role, shape} and writes at most one CQ-01 (code) defect line per rule class, with advisory lines marked (advisory). python -m core.adaptability report plots/ [--sample N] prints the baseline below.
• core/spec_roles.py is the agent's role parser, moved, not copied. The agent and its tests import it from core. It adds data_bullets (pinned to sync_to_postgres.py on every spec), the shape rule with SHAPE_OVERRIDES and MAP_SPECS, and the colour-driving role.
• automation/scripts/adaptability_check.py is the single-file mirror for the workflow runner. core/adaptability_mirror.py generates it, and a test pins the file to the build, keeps it stdlib-only, runs it with python -I where core cannot be imported, and compares its records with the module's on a catalogue sample.
• core/perturbation.py builds the smoke run's tables and the mutated source. It has no nullable-dtype perturbation. core/render_identity.py computes the share of differing pixels against a bound, with a small CLI.
• No workflow, prompt, rubric or score changes. The pipeline does not run any of this until PR 2.

Plan

docs/concepts/adaptable-code.md, "Rollout" item 1, with the decisions of 2026-10-23 from the brief. Only the status line of that file changes here.

Catalogue baseline

uv run python -m core.adaptability report plots/ on main at f2b4629:

Library	Files	Pass	Fail	Excluded
matplotlib	325	8	304	13
seaborn	325	2	310	13
plotly	325	7	305	13
altair	325	0	312	13
plotnine	321	0	308	13
bokeh	325	3	309	13
pygal	325	2	310	13
letsplot	325	0	312	13
all	2596	22	2470	104

Files with a finding, per rule class:

Rule	Class	matplotlib	seaborn	plotly	altair	plotnine	bokeh	pygal	letsplot	All
1	layout	299	281	300	295	269	308	310	288	2350
2	columns	281	281	281	281	277	281	281	281	2244
3	interface	191	120	117	25	136	80	33	110	812
4	column-literal	12	90	12	32	170	53	0	132	501
4	sample-literal (advisory)	11	25	13	12	21	11	3	14	110
4	data-coordinate (advisory)	120	114	117	86	127	91	77	98	830
4	statistic-text (advisory)	8	9	8	7	6	10	12	4	64
4	df-comparison	0	9	0	1	3	1	0	1	15
4	count-literal (advisory)	5	4	7	1	7	1	2	5	32
5	palette	47	51	46	44	47	44	37	43	359
6	constants (advisory)	0	0	0	0	0	0	0	0	0

Specs per shape: 278 table, 19 generated, 9 network, 6 matrix, 13 excluded. That matches the panel's count of 278 tables. All 22 passing files belong to generated specs, where the design asks only for the markers, a clean block and the palette (see "Findings").

Rule 1 fails mostly on missing markers, as the table below shows. Altair is the extreme case: only 43 of its 325 files have a # Plot marker.

Layout finding	Files
# Data marker with no # Plot after it	1,223
A block that does not end by binding df	780
No # Data marker at all	325
A plotting, styling or theme token inside the block	309

Precision

A finding is judged true when the literal or structure ties the code to the catalogue's own data, so that other data would break it or show stale text. Each sample draws random findings, at most one per file, from the catalogue with a fixed seed (--sample N --seed S). The per-class verdicts use the samples taken after the last change to each class.

Class	Rule	Judged after the fixes	True	Precision	Verdict
layout	1	30 (seed 7)	30	100 %	binding
columns	2	30 (seed 4)	30	100 %	binding
interface	3	30 (seed 7)	30	100 %	binding
column-literal	4	35 (seeds 5, 7)	33	94 %	binding
df-comparison	4	15 (every file)	15	100 %	binding
palette	5	30 (seed 5)	28	93 %	binding
sample-literal	4	35 (seeds 4, 7)	30	86 %	advisory
count-literal	4	35 (seeds 3, 7)	26	74 %	advisory
data-coordinate	4	35 (seeds 2, 7)	22	63 %	advisory
statistic-text	4	35 (seeds 3, 7)	12	34 %	advisory
constants	6	none: no catalogue file has a # Constants zone			advisory by decision

The verdict is per class, not per rule. Rule 4 as a whole judges 22 of 30 on the stratified sample below, but two of its classes clear 90 % and stay binding. These fixes came from earlier samples:

• df-comparison went from 14 of 25 to 15 of 15 after inequalities stopped counting. A threshold such as df[RSI] >= 70 or if w > 7 holds for any data.
• palette went from 23 of 30 to 28 of 30. A missing palette now counts only for a required colour role. A strict zip over a palette sliced to len(groups) is allowed. entity is no longer a colour role, since slope charts colour by direction.
• data-coordinate improved from 21 of 30 through four changes. Files with hidden axes are drawings. Padding beside a computed bound is ignored. range= in alt.Scale and scale_size is visual. PIL's draw.text((x, y)) is pixels. The class still stays at 63 %: annotation positions in layout units and domain constants (pH 0 to 14, a rebased-100 line, 45° phase ticks) cannot be separated statically.
• sample-literal stopped counting record field names and one-character values. Its misses are domain vocabulary that the block also spells out: ECG lead names, Venn zones, spectral classes, QR module types.

Rule 1, layout: 30 lines, seed 7 (30 TP)

TP band-basic/letsplot:49                   # Data with no # Plot after it
TP candlestick-basic/pygal                  no # Data marker
TP curve-bias-variance-tradeoff/altair:32   no # Plot
TP curve-dose-response/bokeh:77             block ends with fold_diff = ec50_b / ec50_a
TP curve-dose-response/pygal:58             no # Plot
TP drawdown-basic/matplotlib:78             block ends with a for loop over drawdown
TP ecdf-basic/bokeh:50                      block ends with a ColumnDataSource
TP gauge-basic/seaborn:37                   no # Plot
TP heatmap-calendar/bokeh:80                LinearColorMapper (styling) inside the block
TP heatmap-clustered/bokeh:48               no # Plot
TP heatmap-periodic-table/altair            no # Data marker
TP hexbin-basic/pygal:48                    no # Plot
TP histogram-epidemic/plotnine:92           block ends with a label string
TP histogram-kde/plotly:27                  no # Plot
TP indicator-ichimoku/altair:36             no # Plot
TP ks-test-comparison/letsplot:96           theme token INK_MUTED inside the block
TP line-annotated-events/matplotlib:31      block ends with an events list
TP line-multi/bokeh:29                      no # Plot
TP network-basic/bokeh:143                  block ends with a layout loop
TP network-weighted/altair:24               no # Plot
TP phase-diagram/seaborn:61                 block ends with sns.set_theme(...)
TP point-and-figure-basic/plotnine:123      block ends with a second frame (trend_lines)
TP radar-multi/seaborn:52                   no # Plot
TP scatter-embedding/pygal:47               block ends with centroids
TP scatter-pitch-events/seaborn:78          theme token IMPRINT_PALETTE inside the block
TP shap-waterfall/letsplot:27               no # Plot
TP smith-chart-basic/pygal:76               no # Plot
TP streamgraph-basic/altair:24              no # Plot
TP venn-basic/matplotlib:29                 block ends with a_and_b_and_c = 10
TP violin-swarm/bokeh:31                    no # Plot

Rule 2, columns: 30 lines, seed 4 (30 TP; no catalogue file has role constants yet)

TP bar-feature-importance/altair            feature, importance
TP box-grouped/pygal                        category, subcategory, value
TP campbell-basic/bokeh:36                  speed, frequency_mode_N, engine_order, critical_speeds
TP contour-density/plotly                   x, y
TP contour-density/seaborn                  x, y
TP ecg-twelve-lead/matplotlib:29            time, lead_I ... lead_V
TP energy-level-atomic/bokeh                level, energy
TP funnel-meta-analysis/bokeh:31            effect_size, std_error
TP hexbin-basic/plotnine                    x, y
TP histogram-overlapping/letsplot           values, group
TP indicator-ema/bokeh                      date, close, ema_short, ema_long
TP indicator-ichimoku/plotly:31             date, open, high, low, close, ...
TP lift-curve/altair                        y_true, y_score
TP line-loss-training/altair                epoch, train_loss, val_loss
TP line-parametric/bokeh                    t, x, y
TP line-win-probability/altair              game_time, win_probability
TP line-win-probability/letsplot            game_time, win_probability
TP manhattan-gwas/seaborn:26                chromosome, position, p_value
TP network-bipartite/seaborn:43             source, target
TP qq-basic/plotly:26                       sample
TP raincloud-basic/altair                   category, value
TP roc-curve/bokeh                          fpr, tpr, auc
TP scatter-basic/seaborn:23                 x, y
TP spectrogram-mel/altair                   audio_signal, sample_rate, n_mels
TP spectrogram-mel/plotly:26                audio_signal, sample_rate, n_mels
TP spirometry-flow-volume/plotnine          volume, flow_expiratory, ...
TP star-chart-constellation/letsplot:30     star_id, ra, dec, magnitude, constellation
TP ternary-density/plotly:35                component_a, component_b, component_c
TP violin-split/seaborn:26                  category, value, split_group
TP voronoi-basic/letsplot                   x, y

Rule 3, interface: 30 lines, seed 7 (30 TP)

TP andrews-curves/seaborn:89                curves_df
TP area-basic/plotnine:76                   peak_val
TP bar-3d-categorical/seaborn:63            n_reg
TP bar-basic/plotnine:78                    top_val
TP bar-diverging-likert/letsplot:163        annot_top
TP bar-pareto/plotnine:101                  scale_factor
TP diagnostic-regression-panel/plotnine:208 df_qq_ref
TP ecdf-basic/plotly:61                     p90
TP heatmap-calendar/matplotlib:68           n_weeks
TP heatmap-chromagram/letsplot:113          frame_duration
TP heatmap-chromagram/plotnine:101          time_step
TP heatmap-loss-triangle/matplotlib:79      is_projected
TP histogram-cumulative/seaborn:52          transactions
TP histogram-epidemic/matplotlib:92         cumulative
TP indicator-ichimoku/plotly:164            tenkan_sen
TP line-growth-percentile/letsplot:136      df_patient
TP line-markers/plotnine:90                 anyplot_theme (a theme built in the block)
TP line-multi/plotnine:88                   end_labels
TP line-win-probability/seaborn:98          plays
TP piano-roll-midi/plotnine:164             label_pitches
TP pyramid-basic/plotnine:98                anyplot_theme
TP qq-basic/plotly:76                       line_max
TP radar-basic/pygal:94                     categories
TP shap-waterfall/seaborn:90                bar_lefts
TP spectrum-nmr/plotnine:140                ref_labels
TP spiral-timeseries/matplotlib:82          rev_spacing
TP step-basic/letsplot:74                   label_df
TP stock-event-flags/plotly:97              events
TP swimmer-clinical-timeline/plotly:72      sorted_arms
TP wireframe-3d-basic/pygal:108             col_lines

Rule 4, literals below the block: 30 lines, seed 7, five per class (22 TP)

TP column-literal  datamatrix-basic/plotnine:80          aes(x="x", ...)
TP column-literal  dot-matrix-proportional/plotnine:62   aes(..., y="y")
TP column-literal  heatmap-rainflow/plotnine:93          aes(x="Mean Stress (MPa)")
TP column-literal  span-basic/plotnine:68                aes(y="price")
TP column-literal  stock-event-flags/letsplot:155        aes(y="flag_y") on a frame the block builds
TP sample-literal  bar-spine/plotly:55                   for outcome in ["Active", "On Trial", "Cancelled"]
TP sample-literal  histogram-epidemic/plotnine:155       {"Probable": ...} color map
TP sample-literal  scatter-pitch-events/bokeh:172        row["event_type"] == "pass"
TP sample-literal  scatter-pitch-events/seaborn:156      df["Outcome"] == "Successful"
TP sample-literal  sunburst-basic/bokeh:194              inner_radius="inner_radius", a field the block builds
FP data-coordinate bode-basic/matplotlib:136             phase ticks every 45° (domain ticks)
TP data-coordinate histogram-kde/letsplot:63             geom_rect band at density 0.28
FP data-coordinate line-stock-comparison/plotnine:84     geom_hline(yintercept=100), the rebased baseline
TP data-coordinate phase-diagram-pt/plotnine:183         annotate("text", x=218, y=5e7, label="SOLID")
FP data-coordinate titration-curve/plotnine:143          pH breaks 0..14 (domain scale)
TP statistic-text  curve-oc/pygal:123                    "n=100, c=3" plan typed as text
FP statistic-text  ks-test-comparison/altair:76          "p < 0.001" fallback beside a computed f-string
FP statistic-text  ks-test-comparison/bokeh:129          same
FP statistic-text  ma-differential-expression/pygal:189  "Upregulated (p<0.05)", a threshold
FP statistic-text  skewt-logp-atmospheric/pygal:149      "Mixing Ratio r=10g/kg", chart furniture
TP df-comparison   errorbar-basic/altair:84              df["category"] == "Treatment D"
TP df-comparison   funnel-meta-analysis/seaborn:141      row["study"] == "Martinez 2020"
TP df-comparison   indicator-ema/plotnine:62             df["series"] == "Close Price"
TP df-comparison   line-stepwise/letsplot:63             df["hour"].isin([6, 9, 12, 15, 18, 21])
TP df-comparison   scatter-hr-diagram/seaborn:149        df["Region"] == "Main Sequence"
FP count-literal   ecg-twelve-lead/plotly:164            range(4) rows of the 12-lead layout
TP count-literal   line-basic/plotnine:53                breaks=range(1, 13), the sample's months
TP count-literal   line-filled/plotnine:52               breaks=range(1, 13)
TP count-literal   ridgeline-basic/plotnine:137          breaks=range(12)
TP count-literal   scatter-constellation-diagram/matplotlib:54  range(4), the sample's 16-QAM grid

Rule 5, palette: 30 lines, seed 5 (28 TP)

TP andrews-curves/plotnine:34               IMPRINT 3 of 8
TP area-stacked-confidence/plotly           no IMPRINT list, category colors the stack
TP area-stacked-percent/matplotlib:21       5 of 8
TP bar-3d-categorical/altair:25             5 of 8
TP bar-3d-categorical/letsplot:23           3 of 8
TP bar-grouped/seaborn:28                   3 of 8
TP bar-stacked-labeled/pygal:28             4 of 8
TP bar-stacked/bokeh:33                     4 of 8
TP bar-stacked/pygal:27                     5 of 8
TP biplot-pca/pygal:32                      7 of 8
TP box-notched/matplotlib:81                zip(boxes, IMPRINT, strict=True)
TP bubble-packed/altair:64                  IMPRINT_PALETTE[:5]
TP column-stratigraphic/seaborn:81          zip(age_order, IMPRINT_PALETTE[:5], strict=True)
TP frequency-polygon-basic/plotly           no IMPRINT list, group colors the lines
TP gauge-activity-rings/plotnine:34         3 of 8
FP horizon-basic/altair                     horizon bands use one diverging scheme, not a color per series
TP line-multi/letsplot:24                   3 of 8
TP line-tanabe-sugano/bokeh:141             zip(IMPRINT_PALETTE, terms, strict=True)
TP lollipop-grouped/bokeh:65                zip(series_names, IMPRINT, strict=True)
TP marimekko-basic/matplotlib:65            zip(products, IMPRINT, ..., strict=True)
TP polar-scatter/letsplot:25                3 of 8
TP scatter-embedding/bokeh:33               6 of 8
TP stereonet-equal-area/altair:26           dict(zip(FEATURES, IMPRINT, strict=True))
TP streamgraph-basic/bokeh:29               COLORS 6 of 8
TP swimmer-clinical-timeline/altair:30      ARM_COLORS 2 of 8
TP timeline-basic/plotly:22                 4 of 8
TP treemap-basic/matplotlib:24              5 of 8
TP treemap-basic/plotly:116                 dict(zip(unique_cats_ordered, IMPRINT, strict=True))
TP violin-grouped-swarm/matplotlib:21       COLORS 2 of 8
FP violin-split/seaborn                     a binary split always has two colors

Rule 6 has no catalogue evidence: no file has a # Constants zone yet. Unit tests cover it, and it stays advisory.

Dry run of the perturbation harness

The dry run used the two converted examples, tests/unit/core/fixtures/adaptability/, with matplotlib 3.11.2 and seaborn 0.13.2 in the local venv. For each table it ran head and block, built the table with core.perturbation, swapped the block with smoke_source and rendered both themes. The canvas was checked with core.canvas. The bar-grouped fixture is the design's example, pivoted without fill_value=0.

Table	bar-grouped (matplotlib)	line-basic (seaborn)
Catalogue data	pass, 3200×1800 both themes	pass, 3200×1800 both themes
renamed (one column with spaces, a unit and a dot; categories renamed)	pass	pass
more_categories (5 quarters, 4 lines)	pass	(no categorical column)
fewer_categories (3 quarters, 2 lines)	pass	(no categorical column)
single_group	pass	(no colour role)
fewer_rows	pass (6 of 12 rows)	pass (6 of 60)
more_rows (10×)	pass (120 rows)	pass (600 rows)
scaled (×1,000 + 100,000)	pass	pass
Nine groups, out of palette by design	IndexError: list index out of range

The whole run took 39 s. The stale-label check on text artists belongs to the PR 2 runner and was not run.

Decisions

Open questions of the design (the owner can override any of them):

• # Constants allowlist: scalars, thresholds, ordered domain scales and semantic maps may live there. Rule 6 flags a list, dict or string there that shares a value with the block's sample values, or two data numbers, and is always advisory. An ordered scale that the block also spells out cannot be told from hoisted categories statically, so the smoke run's renamed table is the real test.
• Columns that are not spec roles: they get # role: extra. Any number of them is allowed, they are never checked against the roles, and the agent binds them by name or falls back to the adapter.
• Scenario text that cannot be computed: an event label goes in # Constants with its key, such as a date, and is drawn only when the key lies inside df. The checker does not flag it unless it repeats a sample value, which is advisory.
• df.to_string(buf=...): ban the buf keyword in the agent's validator and keep to_string(). The checker gets no rule, because no catalogue file uses it. The validator change belongs to the agent's readiness PR.
• R loader: base read.csv(..., check.names = FALSE) with today's R image. readr only once the image ships it.
• Persisting the smoke result: yes. impl-merge writes the merged version's smoke result forward next to the static record, with no backfill. PR 2 stages it, and PR 3 lets the mechanical path require it.
• Renaming constants after substitution: no deterministic rename in phase 1. QUARTER = "Region" stays readable, and a slug rename risks colliding with other names.

Implementation choices, decided here and documented in the modules:

• status is pass, fail or excluded. A map spec is never checked, so neither pass nor fail would be true.
• Markers are column-0 comments, # Data or # Plot alone or followed by : or a dash and a hint. The catalogue writes # Data - ... 873 times and # Data — ... 713 times.
• The role comment is # role: <role> with the role name alone, on any line of the assignment. extra is the one non-spec role.
• A block "binds df" by df = ..., by a column or attribute assignment on df, or by an in-place call on it.
• Rule 4's string rules compare whole strings only, so prose that contains a sample value is fine. They skip style keywords, prose words ("Total", "Other"), style words ("center", "dash"), property-like dict keys, one-character values, the title (any string with "anyplot.ai") and output calls.
• Fixed domains are 0 to 1 and -1 to 1 with 20 % headroom, and 0 to 100 and 0 to 360 with 10 %. Neutral values are -1, 0 and 1. CRITICAL_VALUES now lives in core.adaptability, and readiness.py imports it, as it imports MAP_SPECS from core.spec_roles.
• The setup allowlist is per library (SETUP_CALLS). There is one unit test per entry for the head and one for the block.
• The shape rule runs in this order: map specs, overrides, a nodes or edges role, a role named *matrix*, text-only roles as generated, and the table as default. There are 16 overrides, each with a reason. A test asserts that every override changes its spec's shape.
• The colour role is the first categorical or text role whose name is in COLOR_ROLE_NAMES or whose description names a colour encoding. It is tuned for precision, so a pie's category is not detected.
• A colour role whose group count the spec fixes (boolean, or "binary", "two", "pair", "split", "both", "versus" in its description) keeps its groups in every perturbation. Only an unconstrained colour role gets the single-group table, and more_rows never touches the colour column.
• Network specs get no perturbations yet. Their block may bind a nodes table whose ids must follow the perturbed edges, which belongs with network eligibility (from the Copilot review).
• smoke_source takes the perturbation and renames every column constant by its old column name, # role: extra included (rewrite_columns).
• The mirror is generated, not hand-copied: uv run python -m core.adaptability_mirror, with --check for CI-style use.

Findings for the owner

• Statistic-text precision is far below 90 % and no static fix is in sight, at 34 %. Typed thresholds and domain facts look exactly like typed statistics. It stays advisory and is a candidate for removal if the shadow phase confirms that.
• Spec-side roles: about 20 table specs list scalar parameters as required roles, so rule 2 asks for column constants that no tidy table has. Examples are sample_rate and n_mels (spectrogram), lsl, usl and target (histogram-capability), min_value and max_value (gauge-basic), auc (roc-curve), annual totals and elevation (climograph), base_value and final_value (shap-waterfall), and triple_point (phase-diagram-pt). This does not block the parser. It is input for the spec audit the design schedules first.
• Generated specs pass rule 1 to 5 cheaply: 22 files pass today because the design asks only for the markers, a clean block and the palette there. A rule that the block holds parameters only would need a definition of "parameter" first.
• The matrix shape is lenient by design until the misfiled matrix specs are retyped (heatmap-correlation, heatmap-clustered, learning-curve-basic, pdp-basic): rule 2 asks nothing of a matrix.
• No workflow change is needed for this PR.

Test plan

• uv run pytest tests/unit tests/integration -n 4 --dist worksteal: 12,611 passed.
• uv run ruff check . and uv run ruff format --check ..
• uv run --extra typecheck mypy api core.
• uv run python -m tools.changelog check --base origin/main.
• uv run python -m core.adaptability_mirror --check. The mirror test also runs the file with python -I from a directory without core.
• uv run python -m core.adaptability report plots/ for the baseline and the precision samples above.
• Perturbation dry run on both converted examples, in both themes, on canvas.
• The agent suites that import the moved parser (tests/unit/agents/data, runtime, evals, code/test_readiness.py) pass.

[codecov]

Codecov Report

❌ Patch coverage is 94.87871% with 19 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
core/adaptability_mirror.py	79.45%	15 Missing ⚠️
core/perturbation.py	97.66%	4 Missing ⚠️

📢 Thoughts on this report? Let us know!

[MarkusNeusinger]

First review (Gemini 3.8 Flash, routine lane): the run hit its input budget and its answer was cut at the output limit (exit 4), so only part of it arrived. It raised one finding, which I verified and fixed in a6cf527: annotated assignments such as QUARTER: str = "Quarter" # role: category and an annotated IMPRINT were invisible to the checker, so it reported a missing constant and a missing palette. The catalogue baseline in the PR body is unchanged by the fix.

[Copilot]

🟡 Changes recommended

Unresolved checker and source-substitution errors can miss data dependencies or cause compliant implementations to fail smoke checks.

6 open issues

• Track definite bindings across branches and exceptional paths · New
• Preserve role comments in multiline constant rewrites · New
• Exclude constrained colour groups from category mutations · New
• Map renamed extra columns by identity during substitution · New
• Exclude colour keys from numeric and datetime continuations · New
• Preserve companion node tables in network perturbations · New

What changed in this PR

This PR adds the foundation for checking whether Python catalogue implementations can work with replacement data. It shares the agent’s role parser with the checker and introduces smoke-test helpers without enabling them in the pipeline.

Changes:

• Adds static adaptability checks and a generated standalone mirror.
• Adds data perturbations, source substitution, and render comparison helpers.
• Moves role parsing into core, updates consumers, and adds tests and documentation.

File	Description
tests/​unit/​core/​test_spec_roles.py	Tests shared role parsing, shapes, and colour roles.
tests/​unit/​core/​test_render_identity.py	Tests pixel comparisons and CLI results.
tests/​unit/​core/​test_perturbation.py	Tests table mutations and smoke-source generation.
tests/​unit/​core/​test_adaptability.py	Covers checker rules, substitutions, and catalogue reports.
tests/​unit/​core/​test_adaptability_mirror.py	Checks mirror generation, isolation, and parity.
tests/​unit/​core/​fixtures/​adaptability/​line_basic_seaborn.py	Adds an adaptable line-plot fixture.
tests/​unit/​core/​fixtures/​adaptability/​bar_grouped_matplotlib.py	Adds an adaptable grouped-bar fixture.
tests/​unit/​agents/​runtime/​test_stream.py	Uses shared role definitions.
tests/​unit/​agents/​runtime/​test_data_questions.py	Uses the shared role type.
tests/​unit/​agents/​evals/​test_make_fixtures.py	Uses the relocated parser.
tests/​unit/​agents/​data/​test_signs.py	Updates role imports.
tests/​unit/​agents/​data/​test_parse.py	Updates the role-type import.
tests/​unit/​agents/​data/​test_computed.py	Updates the role-type import.
tests/​unit/​agents/​data/​test_bindings.py	Updates role imports and shared test fixtures.
tests/​unit/​agents/​code/​test_readiness.py	Imports map exclusions from core.
docs/​reference/​repository.md	Documents new modules and parser relocation.
docs/​concepts/​agent-network.md	Updates the documented parser location.
docs/​concepts/​adaptable-code.md	Updates implementation status.
core/​spec_roles.py	Centralizes parsing and adds shape and colour-role classification.
core/​render_identity.py	Measures render differences against a configurable bound.
core/​perturbation.py	Builds mutated tables and executable smoke sources.
core/​adaptability_mirror.py	Generates the standalone checker.
changelog.d/​adaptability-checker.md	Records additions and parser relocation.
agents/​main.py	Uses shared roles and map exclusions.
agents/​anyplot/​session_state.py	Uses the shared parser and role type.
agents/​anyplot/​schemas.py	Imports the shared role-name pattern.
agents/​anyplot/​data/​signs.py	Uses the shared role type.
agents/​anyplot/​data/​shape.py	Uses the shared role type.
agents/​anyplot/​data/​computed.py	Uses the shared role type.
agents/​anyplot/​data/​bindings.py	Uses shared role definitions.
agents/​anyplot/​code/​readiness.py	Reuses shared critical values and map exclusions.
agents/​anyplot/​briefs.py	Uses the shared role type.

_{Review effort: Balanced}

[Copilot]

🟡 Changes recommended

Unresolved checker and perturbation defects can produce false passes or reject compliant implementations.

2 open issues

• Handle implicit reads in augmented-assignment targets · New
• Remap matrix role columns after renaming · New

1 previously missed issue

In code that hasn't changed since last review

Exclude fixed colour columns from numeric category scaling

core/​perturbation.py:217

Excluding a fixed colour-group column from varied does not protect it from this numeric-scaling loop. When an axis category is expanded, copied rows with integer group code 1 become group 1.1 while the original groups 0 and 1 remain. A compliant two-group plot therefore receives an invalid third group; with eight numeric groups, this can also exceed the palette limit. Pass the colour column into _more_categories and exclude it from numeric scaling, as _more_rows does. Add a regression combining an axis category with an integer-coded fixed two-group role.

6 resolved since last review

• Preserve companion node tables in network perturbations
• Exclude colour keys from numeric and datetime continuations
• Map renamed extra columns by identity during substitution
• Exclude constrained colour groups from category mutations
• Preserve role comments in multiline constant rewrites
• Track definite bindings across branches and exceptional paths

_{Review effort: Balanced}