feat(rows): say it once, and only when you know it

[lroolle]

Four live reports against v0.31.0, one failure mode: the block under line 1 saying more than it knows.

Row 3 was restating row 2

notice_flash_line assigned pin_key from the first record and never compared anything to it — dead code under a comment promising it "never echoes the pin's own sentence". With one notice collected, the loop's first iteration is the pin, so session start printed:

- 12✕5h left · 5.8%/win   <ledgers>
- budget ~12✕5h left · even 5.8%/win · heading ~85%

Row 3 now starts at record two: it carries the highest-ranked notice the pin is not already saying, or nothing. A calm start is two rows.

A 7d projection needs a day of this window's evidence

Right after a weekly rollover, the learned walk projected the previous window's weekday profile onto a pool 2% spent. Measured on a synthetic 2.4h-old window (21 days history, 30%/day, recent_24h 40):

gate	seven_day_forecast	week_dry_slot	seven_day_pace 8
off	red 95	slot 14	yellow 5 1
on	(silent)	-1	green -1 0

Every input for those verdicts predates the reset — including the L1 recent_24h blend, which describes a trailing day lying on the far side of it.

New SEVEN_DAY_YOUNG_SECS=86400 gates three sites: _profile_walk (exits like a cold start, which silences the forecast notices, the × cells, the budget heading, the accuracy logger, --check and week in one edit), seven_day_pace (falls to the percent-threshold branch), and week_dry_slot's linear fallback. Badges still report real percentages and a genuinely spent young window still goes red on its own 85%+: the guard mutes pace, never facts. The 5h window is untouched.

Pace waits for a fraction of its window

ADVISOR_PACE_MIN_ELAPSED=900 is 5% of a 5h window and 0.15% of a week — an hour into a fresh 7d window, 2% of the pool over 0.6% of the time rendered 3.0✕ in red. strip_tail now gates on length / 20: 5h stays exactly 900s, 7d becomes ~8.4h, the fraction seven_day_elapsed has always called the noise floor. The constant stays for the 5h cap projection, which is what its comment always said it was for.

The fold token counts windows, not cells

The 34-cell grid spans 170h against a 168h period, so ...▯5h✕10 (cells the fold hid) sat on the same row as 12✕5h left (windows to the reset) and the reader had to arbitrate. The token is now ...▯(✕12) — parenthesized, no redundant unit, counting the 5h windows remaining to the true reset from the same instant the budget line prices. Dry tails read ...×(✕12): × the cell, ✕ the operator, one column apart and never the same mark. build_ledger_strip takes the period length instead of a unit label.

The budget voice drops its sigil

! and + mark a notice that interrupts. The week's resting reading interrupts nothing, and a leading - on a dim line reads as a bullet — which is what made rows 2+3 look like a two-item list. The pin reads 12✕5h left · 5.8%/win; --check and week still label the long form with the word budget.

Verification

• 423 bats tests (was 418), shellcheck -S error clean
• both repo fixtures rendered: fold (✕12) matches the budget's 12✕5h left on the same render, two rows not three
• synthetic rollover fixture rendered hermetically: silent at 2.4h old, projections return at 25h

🤖 Generated with Claude Code