do NOT run tests unless specifically instructed to do so or before committing
If tests fail on the same issue twice in a row, STOP and ask the user how to proceed instead of continuing to fix and retry.
ALWAYS confirm that the feature you add / bug you fixed behaves as it should by running the program after your changes (not tests!) and inspecting output comparing it with what user wanted, after confirming by yourself ask user if the program behaves as they intended
After running tests fix all coverage gaps and issues, do not ignore unless specifically instructed to do so
You are NOT done until you install the new version on the phone itself:
`bash ~/.claude/scripts/phone_deploy.sh stronglift_replacement/workout_app`
(it picks a safe versionCode and launches; never a bare `flutter install`).
To drive a whole workout on the phone without touching real data, deploy the
sandbox flavor instead: `... --flavor sandbox`. That is a separate package
(`com.kuhy.workout_app.sandbox`) with its own data, no network, no LAN server,
a red SANDBOX ribbon, and a SANDBOX section in Settings (wipe, "mark today not
done", rest length in seconds). Its trace is `adb logcat -s WorkoutSandbox`.

## NEVER fail silently

A failure that says nothing is a bug. The PC's workout sync did nothing for
weeks because a missing token `return`ed `[]` and real errors only logged at
`INFO` — nobody could tell "nothing to do" from "it broke".

- **Swallowed exceptions must say why, loudly.** Every `except` must either
  re-`raise` or log at `warning`/`error`/`exception`/`critical` with the
  concrete reason. `debug`/`info` do NOT count — an invisible log is a silent
  failure with extra steps. Enforced by `scripts/check_silent_failures.py`
  (pre-commit hook `no-silent-failures`); ruff's `S110` only catches
  `except: pass`, which is the rare case.
- **Guard clauses that abort work must log too.** `if token is None: return`
  is a silent no-op. Say what was missing and what it means.
- **Return something meaningful.** Prefer a typed result (e.g.
  `_manual_push.PushResult(pushed, record_count, reason)`) over a bare
  `None`/`[]`/`False`, so the caller can react instead of guessing. `reason`
  should read like a sentence a human can act on.
- **No escape hatch** — `# noqa` is banned repo-wide. If a failure really is
  benign, log it at `warning` and move on.

## The shutdown hour is derived, never incremented into

`/etc/shutdown-schedule.conf` is recomputed from scratch by
`_shutdown_base.reset_to_base_if_new_day` on each new day: the base plus
every earned term, capped at 23. The numbers are NOT in this repo: they live
in the shared earner registry `earned_time` (kuhyx/utils, `earned_time/`),
which steam-backlog-enforcer reads for gaming time too. Base = 20:00 minus
each earner's cut once its `penalty_from` day arrives (reading, from
2026-10-01 -> 19:00); terms = workout 2h + 1h per further session, LeetCode
1h, reading 1h, plus whatever is registered later.
Anything that pushes the hour with a plain read-add-write and is *not* a term
of that derivation gets wiped by the next reset -- that is how a 00:02 workout
lost its hours on 2026-09-13. A new bonus source is an `Earner` in the
registry, never code here: the reset includes it as a term, and the live pass
(`apply_flat_bonuses_if_new`, stamped `<name>_bonus_date` in
`shutdown_base.json`) picks it up for flat earners. That pass runs the moment
a gate writes its ledger: `earner-bonus.path` (rendered from the registry by
`_earner_units`, re-synced by every `--sync-only` pass) starts
`earner-bonus.service` (`--apply-bonuses`). The 15-minute `workout-sync` is
only the fallback. The reset and the pass hold `_bonus_lock`, since three
processes can run them at once. The base is the registry
constant, not the state file: the file once persisted `base_*_hour`, which
made the code's default dead.

A sick day is a date in the shared `sickdays` pool, read only through
`_sick.py`. It may keep lockers: the workout lock asks
`_sick.stands_down_workout` (keep-aware), while everything that only computes
shutdown (`_sick_floor`, `_shutdown_base`, `_shutdown_target`, the preview
scripts) asks `is_sick` and ignores keep. A sick day's shutdown is absolute:
20:00 plus the GAMING pay of everything credited that day, capped at 23:00
(`earned_time.sick_shutdown_from`, 0.13). The reset writes it, `_sick_floor`
re-derives it on every live pass from the ledgers, and a sick-day credit only
raises to it (`_raise_on_sick_day`), never adds -- its row is already in the
sum. No lift is saved on a sick day.

## A shutdown never blocks submitting a finished task

A credit timed at or after tonight's shutdown moves tonight to
`min(23:00, the credit's OWN time + its reward)` (`_late_credit.late_tonight`):
only the 23:00 ceiling is hard. The anchor is the credit's own time (the run's
end via `_workout_ledger.entry_done_at`, the ledger row via
`_earned.first_credit_time`), never the sync time, so a late upload buys
nothing. It is live-only: the next day's reset re-derives from scratch. This
replaced the first-credit +60 grace floor (gone 2026-10-17); `_grace_floor`
keeps only `first_done_at`, `tonight_minutes` and the lift file, which
`plan_absorb` spends once if a stale one is left.

Flat earners are read from their gate's ledger by `earned_time.done_today`
via `_earned.py`: HMAC-verified `credit` rows matched by the earner's own
rule (LeetCode's `submitted_at`, book-guard's `detail.bonus == "1"` +
`ended_at`), and *fail closed* -- an unreadable ledger or key earns nothing,
logged. The gates stay read-only; none writes the config.

The schedule is in minutes after midnight end to end. `/etc/shutdown-schedule.conf`
holds authoritative `MON_WED_MINUTES` / `THU_SUN_MINUTES` / `MORNING_END_MINUTES`
plus strict-rounded legacy `*_HOUR` copies (evening floor, morning ceil) for old
readers; `read_shutdown_config` prefers `*_MINUTES` and falls back to `*_HOUR` x 60.
`adjust_shutdown_schedule.sh` takes `HH:MM` (or a legacy bare hour) and the
helpers `_earned.hhmm` / `_earned.span` render times and durations, so a
30-minute earner is applied exactly.

## The morning session is the carrot — and the early-bird window is it

`EarlyBirdMixin._is_early_bird_time` no longer reads a clock: it is open
exactly while wake-alarm's signed file grants an exemption (before the
alarm, during the session, until 11:00 once it completed in time). The
05:00–08:30 window and the extended-to-09:00 reward locked someone who got
up at 07:00 and were a second mechanism next to the one that knows whether
the user is up (2026-09-20). The pending marker is banked by the
`wake_alarm_skip` rung; `early-bird-workout-check.timer` (09:30:30, 11:00:30)
re-checks just after the carrot's two possible ends.

The `wake_alarm_skip` rung reads wake-alarm's HMAC-signed
`~/.local/state/wake_alarm/morning_session.json` (`_morning_session.py`), not
Firebase: wake-alarm is the only reader, and the previous channel here pointed
at a path that stopped existing when the repos split and answered "no skip"
for a year without an error. The whole decision is *signature ok, dated today,
now < exempt_until*; the 09:30/11:00 cutoffs live in wake-alarm
(`DOCS-morning-session-pc.md`) and must not be re-derived here. The enforce
path waits up to 30 s for a freshly booted PC's refresher; status paths never
wait, and tests pin the retry to 0.

## Commands

- run: `./run.sh` (locks the screen; never launch on the live display for checks)
- test: `python -m pytest -n auto -p no:sugar -q`
- test-changed: `scripts/test_changed.sh`
- lint: `pre-commit run ruff --all-files`
- coverage: `python -m pytest -n auto -p no:sugar -q`
- coverage-gaps: `coverage-gaps coverage.lcov`