feat(course): record AI numbers lessons 13-21 with t27c test-report and a one-line sed

[gHashTag]

Description

Owner's decision: lessons 13 to 21 of the "AI numbers" course said "Recording pending: it waits on tri test and tri mutate plant from gHashTag/t27#7400". #7400 is blocked, so the nine lessons are now recorded the way term/t27c-gfternary/ was: t27c test-report <spec>, a one-line sed mutation, t27c test-report again showing exactly the test the lesson names failing, then git checkout <spec> and sha256sum (the same hash as before the edit).

Recorded with t27c 0.5.1 and zig 0.16.0 in a Claude Code cloud container (Linux x86_64), from a clean worktree of gHashTag/t27 at 72f793357. Every lesson claim was first reproduced by hand, then recorded:

Lesson	Spec	Mutation (one sed line)	Fails	Matches the lesson text
13 a-sign-is-one-bit	gft_smul.t27	sign always 0	m2	yes
14 opposite-signs-subtract	gft_sadd.t27	magsub -> magadd in the branch a2 takes	a2	yes
15 multiply-then-add	gft_signed_mac.t27	sign XOR -> OR	pp	yes
16 a-bend-at-zero	gft_relu.t27	negative inputs pass through	negz	yes
17 two-to-the-x	gft_exp2.t27	minus sign of k dropped	em1	yes
18 pick-the-largest	gft_argmax4.t27	> -> >= for two positive scores	tie_low	yes
19 a-loss-in-bits	gft_nll.t27	neg returns 65536 for 0	perfect	yes
20 one-step-downhill	gft_sgd_step.t27	smul(eta, g) -> smul(g, g)	ascend	yes
21 xor-needs-a-bend	gft_xornet.t27	relu removed from h1	x00	yes

The lesson-18 side claim was checked too: reversing the comparison of two negative scores fails no test.

Related Issue

Related to gHashTag/t27#7400 (still blocked; no longer a dependency of these lessons). Related to gHashTag/t27#7370 (t27c test-report exits 0 on a failing test, so verdicts are read from the report text).

Changes Made

• Feature: apps/website/scripts/t27c-mutation-casts.mjs -- recorder (reuses bareEnv from tri-help-casts.mjs). Staged: prompt and typing. Real: every byte the commands printed, at the time they printed it. It refuses to write a recording whose baseline is not all green, whose mutant does not fail exactly the named test, whose diff is not one line, or whose hash differs after git checkout; it requires a clean t27 checkout before and after.
• Recordings: 9 x public/term/t27c-gft-*/ -- session.cast (asciicast v2), meta.json (no pending; gif: null; desc names container, t27c, zig, t27 commit), index.html, card.png; public/term/index.html relisted (count stays 104).
• Docs: the nine lesson texts in specs/course/ai-numbers.t27 and i18n/ai-numbers.ru.json describe the recording that exists; gallery hooks in specs/widgets/gallery.t27; a dated update paragraph (EN + RU) in the two-courses blog post.
• Tests: scripts/t27c-mutation-casts.test.mjs, wired into npm run test:course.
• Regenerated with the repo's own generators: widgets-from-spec.mjs, course-from-spec.mjs, course-pages.mjs (18 lesson cards redrawn).

GIF

No GIF encoder was used, so gif is null and there is no session.gif. widgets-from-spec.mjs then uses the card in the gallery markdown (gif ?? image); no contract requires a GIF.

Test Results

All 65 steps of the checks job of .github/workflows/website-checks.yml were run locally, one command at a time: 63 pass. Two fail, the same way on origin/main:

• Queen viewport matrix and Queen touch contract: sectors=0 / page never became ready - sectors 0/6. On origin/main a0d149e (a clean worktree, same container) both fail identically: Queen viewport matrix with the same 18 size/view combinations, Queen touch contract with sectors 0/6. Logs: local runner, steps 57 and 63.

Breaking Changes

• No - backward compatible; URLs of the nine recordings are unchanged.

TOXIC VERDICT

What Works

• Nine real recordings; each shows the baseline green, the one named test red, and the spec hash restored. The recorder refuses anything else, and a test re-checks the committed casts.

What Doesn't Work

• No GIFs. t27c test-report exits 0 on a failure (t27#7370), so the exit codes in meta.json are all 0 and the verdict lives in the report text.

Tech Debt

• The recorder's lesson table (lines to show, sed, failing test) lives in the script, next to the lesson text that states the same facts; the test ties the committed casts to that table.

Known Issues

• Recorded in a cloud container, not on the t27c lab; the exact commit t27c 0.5.1 was built from is not recorded.
• tri-test-now-runs-the-tests still says the lessons wait on #7400 (left as written on its date).

Overall Assessment: APPROVE once the Queen checks are confirmed green in CI. Self-Score: 8/10

TECH TREE Options

1. GIFs for the nine -- render session.gif from the casts so the gallery shows motion. Low effort.
2. Move the mutation table into a .t27 spec read by the recorder and the lesson text. Medium effort; one source for both.
3. Land t27#7370 so t27c test-report exits non-zero on a failure, then re-record. Recommended: it removes the one caveat these recordings carry.

Checklist

• Self-review of my own code
• Tested locally (see above)
• No secrets; home directories are not shown in any recording (redacted: [])

{
  "version": 1,
  "head_sha": "70b61bd4515ab69683a65f2ba9db2eae586c3660",
  "summary": "Lessons 13 to 21 of the AI numbers course now open real recordings instead of placeholders: each runs t27c test-report, plants one bug with a one-line sed, shows exactly the named test failing, and restores the spec with git checkout; all nine lesson claims reproduced.",
  "changes": [
    "apps/website/scripts/t27c-mutation-casts.mjs: a recorder in the shape of tri-help-casts.mjs (its bareEnv): staged prompt and typing, real output at the time it printed; refuses a recording whose baseline is not all green, whose mutant does not fail exactly the named test, whose diff is not one line, or whose spec hash differs after git checkout; writes session.cast, meta.json, index.html, card.png and relists public/term/index.html",
    "public/term/t27c-gft-{smul,sadd,signed-mac,relu,exp2,argmax4,nll,sgd-step,xornet}: real asciicast v2 recordings replace the placeholders; meta.json has no pending field, gif is null because no GIF encoder was used, desc names t27c 0.5.1, zig 0.16.0, a Claude Code cloud container and t27 72f793357",
    "specs/course/ai-numbers.t27 and i18n/ai-numbers.ru.json: the nine lesson texts no longer say the recording waits on t27#7400; they describe t27c test-report, the sed line, the second report, git checkout and sha256sum, keeping every other fact",
    "specs/widgets/gallery.t27: the nine widget hooks describe the recordings instead of the placeholder",
    "Regenerated with the repo's generators: widgets-from-spec, course-from-spec, course-pages (18 lesson cards redrawn, EN and RU)",
    "scripts/t27c-mutation-casts.test.mjs, added to npm run test:course: the committed casts show exactly the named test failing and the spec hash restored",
    "src/data/blog/bodies/two-courses-twenty-seven-lessons-each.ts: a dated update paragraph (EN and RU) says the recordings exist now and how they were made; the original paragraph is kept"
  ],
  "tests": [
    {
      "command": "t27c test-report specs/ternary/<spec>.t27, then sed -i <mutation>, then t27c test-report again, for all nine specs, by hand before recording",
      "status": "passed",
      "result": "baseline all pass, none vacuous; the mutants fail exactly m2, a2, pp, negz, em1, tie_low, perfect, ascend, x00",
      "evidence": "t27c 0.5.1, zig 0.16.0, clean worktree of gHashTag/t27 at 72f793357; the same runs are in public/term/t27c-gft-*/session.cast"
    },
    {
      "command": "node scripts/t27c-mutation-casts.mjs --t27 <clean t27 worktree> --t27c <t27c> --zig <zig>",
      "status": "passed",
      "result": "wrote 9 recordings, none refused; the t27 checkout was clean before and after every one",
      "evidence": "public/term/t27c-gft-*/meta.json (exit codes all 0, recorded 2026-10-09)"
    },
    {
      "command": "npm run test:course",
      "status": "passed",
      "result": "40 of 40 tests pass, 13 of them in the new t27c-mutation-casts.test.mjs",
      "evidence": "local run at this head"
    },
    {
      "command": "every step of .github/workflows/website-checks.yml (checks job), one command at a time",
      "status": "failed",
      "result": "63 of 65 steps pass; Queen viewport matrix and Queen touch contract fail with sectors=0 because the Queen board's live data is not reachable from this container",
      "evidence": "On origin/main a0d149e7 (a clean worktree, same container) both fail identically: Queen viewport matrix with the same 18 size/view combinations, Queen touch contract with sectors 0/6. Logs: local runner, steps 57 and 63."
    },
    {
      "command": "npm run check:course && npm run check:course-pages && npm run check:widgets && npm run check:widget-pages",
      "status": "passed",
      "result": "all up to date after regeneration; 392 pages, 392 cards, 194 widgets",
      "evidence": "local run at this head; part of the 63 passing steps"
    }
  ],
  "limitations": [
    "gHashTag/t27#7400 stays blocked; the recordings use t27c test-report, which still exits 0 when a test fails (t27#7370), so each verdict is read from the report text and the recorder checks that text",
    "The recordings were made in a Claude Code cloud container, not on the t27c lab machine that made the earlier casts; t27c 0.5.1 was built from a recent tree, its exact commit is not recorded",
    "No GIF: gif is null and the gallery links the card; there is no session.gif",
    "The blog post tri-test-now-runs-the-tests still says the nine lessons wait on #7400; it is left as written on its date",
    "Queen viewport matrix and Queen touch contract were not green locally, on this branch or on origin/main; they need the Queen's live data"
  ],
  "tags": [
    "courses",
    "recordings",
    "mutation",
    "ternary"
  ],
  "blog": {
    "title": "Nine lessons get their recordings without waiting for tri mutate plant",
    "summary": "Lessons 13 to 21 of the AI numbers course were placeholders waiting on a blocked PR; they now show t27c test-report passing, one planted bug failing exactly one named test, and git restoring the spec byte for byte.",
    "outline": [
      "The problem: nine lessons said their recording waited on tri test and tri mutate plant from t27#7400, and #7400 is blocked, so the lessons showed no run at all.",
      "How it works: the recordings use t27c test-report directly, one sed line per lesson and git checkout, and a recorder that refuses any run whose mutant does not fail exactly the named test.",
      "The boundary: t27c test-report still exits 0 on a failure, the recordings come from a cloud container rather than the lab, and there is no GIF."
    ],
    "cast": "term/t27c-gft-relu/session.cast",
    "reproduce": [
      "t27c test-report specs/ternary/gft_relu.t27",
      "sed -i 's/== 1) { return 0; }/== 1) { return x; }/' specs/ternary/gft_relu.t27 && t27c test-report specs/ternary/gft_relu.t27",
      "git checkout specs/ternary/gft_relu.t27 && sha256sum specs/ternary/gft_relu.t27"
    ]
  }
}

🤖 Generated with Claude Code

https://claude.ai/code/session_01ScHVrGSr6zZUdwkdC8DR9k

---

Generated by Claude Code

[github-actions]

X Brain Health Check

Score: 100.0/100
Status: 🟢 HEALTHY
Threshold: 80/100

X Brain is above merge threshold

[github-actions]

🧪 Brain Stress Test Results

Tests: 261/261 passed, 0 leaked
Status: 🟢 PASS
Gate: every test in zig build test-brain-stress passes and none leaks