Add theme-aware figures: restyle fig-catalog-status, add row-basis figure

.github/workflows/ci.yml

@@ -127,3 +127,27 @@ jobs:
         run: |
           chock check
           chock sync --repo . --check
+
+  # A stale generated figure fails the build: regenerate every docs/figures/*.svg and diff against
+  # what is committed. make_row_basis.py and make_social_card.py both shell out to the README's
+  # own quickstart block (see tools/quickstart_block.py) to draw their numbers from a real,
+  # freshly produced report.json, so context-report itself must be installed first.
+  figures:
+    name: Generated figures are not stale
+    runs-on: ubuntu-latest
+    # Reads the checkout; writes nothing (the regenerated files are diffed, never pushed).
+    permissions:
+      contents: read
+    steps:
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7
+        with:
+          persist-credentials: false
+      - uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97  # v7
+        with:
+          python-version: "3.12"
+      - run: pip install --require-hashes -r requirements/ci.txt
+      - run: pip install --no-deps -e .
+      - name: Regenerate figures
+        run: cd docs/figures && for g in make_*.py; do python "$g"; done
+      - name: Fail on drift
+        run: git diff --exit-code -- docs/figures

README.md

@@ -95,6 +95,14 @@ it). Three real rows from the `report.json` this exact block just produced:
 relative path `./my-plugin` you passed, so the hook only starts from the one cwd where that path
 still points at the plugin — the exact failure mode the measurement below calls out.
 
+The full statement has 11 rows, not 3 — every row's basis, from this exact quickstart block run
+fresh:
+
+<picture>
+  <source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/open-coder-ai/context-report/main/docs/figures/fig-row-basis-dark.svg">
+  <img alt="Of 11 rows in the report.json this quickstart block produces, 3 are recomputed from the artifact, 1 is the author's unverified claim, and 7 have not been measured this run — shown as absence of evidence, not a weak score." src="https://raw.githubusercontent.com/open-coder-ai/context-report/main/docs/figures/fig-row-basis-light.svg" width="760">
+</picture>
+
 Beyond one statement at a time, `context-report run` drives a whole manifest — subjects × models ×
 tasks — and `context-report compare` puts every run of that manifest side by side; see
 [`docs/cli.md`](docs/cli.md) for the manifest schema, `--dry-run`/`--n`/`--resume`, and which
@@ -143,7 +151,10 @@ skills. Three findings from that run:
 `FAILED, 0 of 4`, exit 126. Every hook that runs, across both samples, allows on malformed input.
 See [§5.2](paper/context-report.md#52-top-n-catalog-plugins).
 
-![Eighteen plugins by four measured attributes](paper/figures/fig-catalog-status.svg)
+<picture>
+  <source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/open-coder-ai/context-report/main/docs/figures/fig-catalog-status-dark.svg">
+  <img alt="Grid of 18 catalog plugins by 4 measured attributes: three plugins fail reachability outright, every plugin that measures malformed-input handling passes it, and not-applicable cells are shown as a neutral dash, never a low score." src="https://raw.githubusercontent.com/open-coder-ai/context-report/main/docs/figures/fig-catalog-status-light.svg" width="760">
+</picture>
 
 **Cost spans two orders of magnitude.** Hooks that shell out to `npx` cost 916.8–941.5 ms p50; a
 local script costs 7.3–53.9 ms. Context weight varies about a hundredfold across the sample,
@@ -244,6 +255,11 @@ merge either way.
 
 ## Part of open-coder-ai
 
+<picture>
+  <source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/open-coder-ai/context-report/main/docs/figures/family-dark.svg">
+  <img alt="The open-coder-ai family: agentseam is the foundation, chock sits on it, chock-catalog feeds chock and generates the four plugin repositories, chock-threat-intel feeds the catalog, and context-report runs as a verification arm measuring all four." src="https://raw.githubusercontent.com/open-coder-ai/context-report/main/docs/figures/family-light.svg" width="800">
+</picture>
+
 | | |
 |---|---|
 | [agentseam](https://github.com/open-coder-ai/agentseam) | the primitives — one handler API and a verified capability matrix across 16 agents |