Standalone script that renders the v0.0.7 result panels as 16:9 slides
sized and styled for talk presentation. Three slide types:

- delta     — "Does Vera beat Python / TypeScript?" headline chart
- tiers     — Flagship and Sonnet tier comparisons side-by-side
- all-modes — all 6 models × 4 modes in a single grouped-bar panel

Specialised, not general. The v0.0.7 model lineup (K2.5 in flagship,
K2 Turbo in sonnet) is hard-coded because the live plot_results.MODELS
registry now reflects the post-K2.6 migration (PR #69). Reuses palette
+ extract_data() from plot_results.py so slide numbers match the README
chart by construction.

Typography is bumped roughly 3× from the documentation-chart sizes so
the slide reads from the back of a room. Slide canvas is 16×9 in at
dpi=180 (2880×1620 px), with a tunable --background flag offering four
light-theme options (paper / white / cream / light-grey). Dark mode is
deliberately out of scope — would require cascading text-colour
inversion that the current talk's design doesn't need.

Output handling:
- PNGs default to /tmp/ because they're talk-prep ephemera that belong
  in the speaker's deck rather than the repo
- assets/vera-bench_slide_*.png is gitignored for the case where
  someone outputs to assets/ for preview — the canonical artefact is
  the script itself; regeneration is cheap
- Brief section added to scripts/README.md flagging the v0.0.7 pin so
  future-me doesn't try to use this for a different release

Verified:
- ruff check + ruff format --check clean
- All three slides render cleanly on default paper background
- Numbers cross-check against the v0.0.7 documentation chart (each of
  6 models × 4 modes = 24 cells)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Two CodeRabbit findings, both valid against current code:

1. --version accepted any string but the model lineup was always patched
   to MODELS_V_0_0_7. Passing e.g. --version 0.0.9 silently produced a
   chart with v0.0.7-era labels and v0.0.9 numbers — exactly the kind
   of label/data mis-mapping that would corrupt a slide deck without
   anyone noticing. Verified locally: --version 0.0.9 produced a PNG
   pre-fix.

   Fix: argparse choices=["0.0.7"] plus help text explaining why the
   lineup is pinned and what future-extension would require. Now
   --version 0.0.9 errors out loud:

       error: argument --version: invalid choice: '0.0.9'
       (choose from 0.0.7)

2. Six internal helpers lacked return type annotations. Added them:

   - _patch_models_for_slide() -> tuple[ModuleType, list[ModelSpec]]
   - _load_v0_0_7_data() -> tuple[dict[str, dict[str, int]],
                                  dict[str, dict[str, int]]]
   - _slide_rcparams() -> None
   - _style_ax() -> None
   - _draw_tier_panel() -> None
   - main() -> None

   New import: `from types import ModuleType` for the patch helper's
   precise return type. Dict type spellings match the pattern in
   plot_results.py (no type alias introduced; inline tuple types
   stay consistent with the existing extract_data signature).

   Argument-level annotations on _style_ax(ax) and _draw_tier_panel(ax)
   left unannotated — out of scope (CR asked specifically for return
   types) and matplotlib's Axes typing has its own complications worth
   handling separately if/when we type the ax parameters globally.

Verified:
- ruff check + ruff format --check clean
- Default invocation still produces all three slides correctly
- --version 0.0.9 now errors at argparse rather than silently
  producing a mis-labeled chart

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

…panel

CodeRabbit pushed back on my earlier deferral of typing `ax`
parameters, correctly pointing out that CLAUDE.md says "type hints
everywhere" without an exemption for matplotlib types. My earlier
"matplotlib Axes typing has its own complications" framing was
sloppy — `from matplotlib.axes import Axes` is a one-line import,
not actually complicated.

Fix:
- Added `from matplotlib.axes import Axes` import
- `_style_ax(ax: Axes) -> None`
- `_draw_tier_panel(ax: Axes, ...) -> None`

Note for follow-up: plot_results.py has the same untyped-ax pattern
across three functions (_style_ax, plot_tier, plot_all_modes).
CodeRabbit's finding would apply there too, but that's a different
file's surface area and out of scope for this PR.

Verified: ruff clean, default invocation produces all three slides
correctly, no behaviour change.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>