Skill card t27/hot-reload-steward: a standing steward for the tri hot paths, reporting on the hot-reload ledger #9084

Refs #8737

The hot-reload loop of 10-11 October (ledger #8737, epic #8095) made every place t27 builds and waits fast, and put each speed-up behind one tri hot word (#8872, #8953):

The session that drove it was archived on the owner's word. The loop needs a standing owner, so it goes to the Queen as a profile: a skill card in t27's own language, plus the SKILL.md it points at. T holds it by rule 1. M (Mu, Metrics / Telemetry, "no regression > 5%") is the domain lead by rule 2, because the job is keeping measured budgets from regressing.

Owner-gated, and not this card's to decide: #8829 (H2, the zig relink), #8900 (the C-header gate cache), #8754 (the nextpnr routeClock patch), #8794 (/bench approve), and the bench poll interval. The card reports on them; it never acts on them.

User Scenarios

1. The Queen's scheduler fires skill/t27/hot-reload-steward.run. A bee opens the steward's SKILL.md and runs it in its Railway container on master. It posts the times against the budgets as one comment on Hot reload: real-time builds everywhere t27 waits -- measured costs, competitors, plan H1-H5 (#8095) #8737.
2. A cache regresses: a warm frontier takes over 5000 ms, or a spec in the card's SPECS stops passing. The bee files one Queen-shaped issue that names the commit and the number, with Refs #8737.
3. A reader of the M agent card sees that M holds the hot-reload skill, and the alphabet table agrees.

Requirements

• FR-001: specs/skills/t27-hot-reload-steward.t27 follows specs/skills/README.md, modelled on specs/skills/t27-providers-steward.t27.
  • Constants: module skill_t27_hot_reload_steward;, KIND "skill", ID "t27/hot-reload-steward", NAME "hot-reload-steward", REPO "t27", SOURCE "SKILL.md", a one-line SUMMARY_EN, COMMAND "", TAGS ("hot-reload", "build", "cache", "steward"), ENABLED true, TIMEOUT_MIN 30.
  • SPECS is these seven: specs/tri/hot/hot.t27, specs/verified/digest_cache.t27, specs/verified/suite_jobs.t27, specs/verified/gen_hash_cache.t27, specs/verified/bitstream_reuse.t27, specs/verified/watch.t27, specs/verified/bench_agent.t27.
  • A test block asserts every constant, including SPECS.len == 7. ASCII only.
• FR-002: the steward's SKILL.md (new, under the .claude/skills/ folder named in the Boundary) has four sections, in this order:
  1. Map. The seven SPECS, the tri hot words, the ledger Hot reload: real-time builds everywhere t27 waits -- measured costs, competitors, plan H1-H5 (#8095) #8737, the epic EPIC: verdict reuse in the language -- rebuild only what is not yet verified #8095, and the post at t27.ai/blog/from-a-minute-to-a-second/. Point at the files; do not restate them.
  2. One run, on master, in the container.
     • t27c test-report on each of the seven SPECS.
     • t27c frontier twice, timing the second run against the 5000 ms budget.
     • tri hot instead, if it is on PATH.
     • The silicon reuse hit needs an openXC7 toolchain (T27_OPENXC7). Where that is unset, say "silicon: not measured". Never write a pass for it.
     • Report the machine's load next to every time.
  3. Report. One comment on Hot reload: real-time builds everywhere t27 waits -- measured costs, competitors, plan H1-H5 (#8095) #8737 per run, giving the commit, the commands, each time against its budget, and the pass counts. One new Queen-shaped issue per regression (## Boundary last, ## User Scenarios, ## Requirements, ## Success Criteria, Refs #8737).
  4. Never.
     • Never merge, approve or auto-merge.
     • Never touch a board, a JTAG cable or flash, and never post /bench approve.
     • Never edit bootstrap/src/test_report.rs, bootstrap/src/cc_gate.rs or tools/policy/foreign-exceptions.txt.
     • Never hand-write code in any language other than t27.
     • Never run builds on the owner's Mac.
     • Never claim a speed-up without the load it was measured at.
• FR-003: the holders.
  • The T card in specs/agents/ gains "t27/hot-reload-steward" in SKILLS: sorted, array length plus one. Its SKILLS_NOTE count is updated to match.
  • The M card gains it the same way. Its SKILLS_NOTE says 6 skill card(s).
  • docs/agents/AGENTS_ALPHABET.md: the M row of the table under "Who holds a tool, and who holds a skill" counts 6 skills and lists t27/hot-reload-steward.
• FR-004: a docs/now/ entry saying what was added. It also says the scheduler sees the card only after gHashTag/BrowserOS trios/agent-server/scripts/sync-t27-specs.sh re-vendors specs/skills and the agent server deploys. Do not claim it is scheduled before that.

Success Criteria

• t27c typecheck specs/skills/t27-hot-reload-steward.t27 prints Typecheck OK (0 errors, 0 warnings)
• t27c test-report specs/skills/t27-hot-reload-steward.t27 prints FAIL 0
• grep -c "SPECS.len == 7" specs/skills/t27-hot-reload-steward.t27 prints 1
• grep -c "t27/hot-reload-steward" specs/agents/t.t27 prints 1
• grep -c "t27/hot-reload-steward" specs/agents/m.t27 prints 1
• grep -c "t27/hot-reload-steward" docs/agents/AGENTS_ALPHABET.md prints at least 1
• grep -c "8737" .claude/skills/hot-reload-steward/SKILL.md prints at least 1
• grep -c "not measured" .claude/skills/hot-reload-steward/SKILL.md prints at least 1

Not in this card's boundary, because other tasks hold those files (#8959 the t27b expectations ledger, #8552 the seals directory): the new spec's seal and its t27b ledger row. Whoever lands the PR adds both in the same PR (t27c seal --save, and the row {"path": "specs/skills/t27-hot-reload-steward.t27", "t27b": "pass"} in the t27b expectations ledger), or master's t27b-native-ratchet goes red as UNLISTED.

Boundary

specs/skills/t27-hot-reload-steward.t27
.claude/skills/hot-reload-steward/SKILL.md
specs/agents/t.t27
specs/agents/m.t27
docs/agents/AGENTS_ALPHABET.md
docs/now/2026-10-11-hot-reload-steward-skill.md