t27f: scene renderers for the 12 archetypes and the still anomaly gate (R3)

[gHashTag]

Closes #10571. Part of epic #10560 (R3).

What this does

reel_factory.t27 (PR #10566) plans one of 12 scene archetypes per lesson, but no code drew a plan. This PR gives each archetype a renderer that turns the lesson's own content into a still, and adds the gate that looks at the pictures.

• Shared parts (built on the reel_still work, PR claw: reel_still.t27 -- the TRI CLAW /mcp tool that renders one particle still #10548): scene_raster.t27 (integer raster: rgb, discs, rings, segments, the bitmap font), scene_lex.t27 (lesson lines, bytes and keyword counts), scene_bg.t27 (lesson-seeded backdrop).
• 12 renderers, one .t27 each, <name>_px(t, acc, n, x, y) on a 360 x 640 canvas: scene_code, scene_graph, scene_tests, scene_trit, scene_loop, scene_branch, scene_data, scene_links, scene_guard, scene_title, scene_task, scene_map. scene_still.pixel dispatches by archetype and applies the plan's accent, motion and layout.
• still_gate.t27: renders the 27 course-0 stills (108 x 192 luma), scores all 351 pairs with an integer block SSIM (12 x 12 blocks, i64, the same number on every machine) and refuses the course when the most similar pair is above SIMILAR_MAX_PERMILLE (850).

Result

Measured on the Railway lab, never on the Mac: the most similar pair of course 0 is lessons 8 and 20 at 607 permille, 243 under the bound. The gate's test pins that pair, so a renderer change that moves it must move the test. Negative controls: a still against itself scores 1000 and is refused; each renderer's tests check that another lesson draws a different picture and that the same lesson twice is one picture.

Verification

• t27c test-report on all 17 new specs: every test passes, 0 vacuous passes.
• t27c seal --save then --verify: all hashes match; the 17 seals are in this PR.
• t27b: none of the 17 specs is rejected (t27b corpus specs/t27f --list on the lab shows no UNSUPPORTED for them). The lab is x86 and the JIT runs on arm64 only, so the pass verdict itself is the arm64 CI's. Ledger: 17 pass rows, max_not_pass unchanged at 134. @mod is replaced by fmod (t27b rejects @mod), and no spec has a printing main.
• Skill t27f content-factory row updated.

Not claimed

SSIM on luma sees layout and contrast, not meaning: two stills with different words in the same boxes would pass. lesson_reel still draws token_floor (R4, next). No render output is committed (nothing under out-reels/).

Stack

The branch is stacked on #10566 (reel_factory) and #10548 (reel_still, draft, based on #10508): their commits show in the diff until they merge. The new work is specs/t27f/scene_*.t27, specs/t27f/still_gate.t27, their seals, 17 ledger rows and the skill row.

🤖 Generated with Claude Code