fix: cascades follow-ups (power-off ops mirror, AST coverage, manual-slider race)

[hughobrien]

Summary

Four follow-up fixes for gaps I flagged after PR #240 landed.

1. Power-off → timer-clear at the ops layer

Probed the office device 2026-05-12 with the night timer (the original Test 4 used turbo, whose short configured duration self-cleared mid-test and made the result inconclusive). Firmware DOES clear 0x0007 when 0x0001 goes 1→0. Encoded in pkg/breezy/ops.go::Power(false) so the daemon's cache stays coherent for ALL callers — UI handler, /v1 JSON, HomeKit, scheduler — not just the /power UI handler, which used to do its own explicit SetTimer(off) and is now simplified.

2. AST-walking coverage test

TestCascadeTable_AllWritableSignalsCovered no longer relies on a hand-maintained whitelist. It now parses the templates package's .go files, walks every clickAction(...) call, and asserts each literal 2nd-argument signal name appears in the cascades map. Mutation-verified: changing "heater" to "unregisteredSig" in controls_block.templ produces a precise file:line error.

3. Manual slider phantom-POST race

The manual slider drag posts dragged value Playwright test flaked 40-60% under heavy load. Stack-trace probing (2026-05-12) pointed at u.rx in datastar-1.0.1.min.js:4:3121: datastar's data-bind plugin fires the change handler synchronously during its signal→input syncs — both initial and on every SSE-pushed signal update. The synthetic events have isTrusted=false and the input's current value, producing no-op {manual:50} POSTs that pre-empted the test's deliberate drag-to-75 assertion.

Fix: replace data-bind with the explicit split that doesn't synthesize change events:

• data-effect="el.value = $_manualPct.X" — signal → input.value (direct property assignment, no DOM event)
• data-on:input="$_manualPct.X = parseInt(evt.target.value, 10)" — input → signal, live drag mirror
• data-on:change={ manualChangeExpr } — POST on release (NO __debounce.200ms — that 200ms window let SSE pushes reset el.value via data-effect between dispatch and handler-fire, making the handler read the server value instead of the dragged value)

15+ full-suite runs under heavy load (concurrent go test -race): all green.

4. Spec updated

Test 4 result settled. Layer B server-side mirrors table reorganized around ops-layer placement now that both speed and power use the same pattern.

Test plan

• just ci green
• 28/28 Playwright tests pass
• New pkg/breezy.TestPowerOff_ClearsTimer and TestPowerOn_NoTimerClear
• Coverage test mutation-verified (changing a signal name in a click handler triggers a precise CI failure)
• Live deployment at breezy.wg now serves the cascade markup (effPower function present, const __next patterns in SSE-pushed cards)

🤖 Generated with Claude Code