Delete the 1 Hz step estimate; rebuild movement minutes on measured evidence - #35

A user reported 2,645 steps on a day they took under 400. Diagnosed on their
real DB: the number was 23 "active minutes" x an assumed 115 spm cadence. Both
halves of that conversion are invalid at 1 Hz and neither is fixable by
retuning, so the conversion is removed rather than recalibrated.

  * Cadence is NOT IDENTIFIABLE. Gait is 1.4-2.3 Hz (Straczkiewicz 2023,
    doi:10.1038/s41746-022-00745-z); at 1 Hz every fundamental is sub-Nyquist
    and 80/100/140/160 spm alias to the same 0.333 Hz. No published step
    detector exists below 10 Hz.
  * The minutes were never specifically ambulation. At the wrist, arm work
    out-accelerates walking (stirring ~104 mg, chopping ~139 mg vs walking
    ~66 mg ENMO), so a movement threshold cannot isolate gait even at full
    rate: wrist devices emit 22-27 false steps/min during dishes, reaching and
    driving (O'Connell 2017, doi:10.1371/journal.pone.0169616) while detecting
    slow walking at sensitivity 0.05. The errors have OPPOSITE sign, so no
    gain constant corrects both.

Confirmed against that DB's own ground truth: in the single window where the
100 Hz pedometer and 1 Hz data overlap, HR ran 95->108 with dynAmp 0.31-0.40 g
and the REAL count was 11 steps in 3.1 min (3.5 spm) where the estimator would
have assigned ~115 spm.

`dailyStepEstimate` -> `dailyActiveMinutes`, returning `DailyMovementEstimate`:
active minutes only, no steps, no cadence, no range. A contract test asserts
the serialised surface can never regain a step or cadence key.

Everything below was PROVEN against 4 days of real 1 Hz substrate first:

  * HR GATE DELETED. `restingHr + 8 bpm` changed the answer by exactly ZERO
    minutes on every day tested. At RHR ~62 it sits at ~6% of heart-rate
    reserve (below every ACSM band) and 73-100% of covered minutes already
    cleared it. It failed in the wrong direction too: PPG HR is least reliable
    during the motion being gated, so a dropout deleted minutes the
    accelerometer measured fine. The API no longer accepts HR at all.
  * x3 CEILING DELETED. Rejected ZERO minutes on all 4 days with 0.42-0.55 g
    of headroom, and cannot fire on artifacts (a 3 s knock averages ~0.23 g,
    below the FLOOR). The only thing it could ever exclude was real exertion.
  * FLOOR MUST BE FROZEN. It is derived from the signal it thresholds, so a
    continuously recomputed floor cancels the trend it exists to report:
    scaling a real day's dynAmp gave 37 active minutes at 1x, 1.5x, 2x AND 3x
    when recomputed, versus 23 -> 254 frozen. Adds
    `enrollmentDaysForFrozenFloor` and `shouldRefreezeFloor`.

REFUTED and deliberately not built: a sleep-anchored floor (CV 138.6% across
days vs 9.3% for the current estimator, and on one night it landed above the
entire day's range, which would report zero); accel autocalibration (offset
and uniform gain cancel exactly through the high-pass and the floor
normalisation -- +5% gain moves the gate decision by 0.0000; only anisotropic
gain survives at ~1-3%); gravity/forearm orientation (it solved the ambulation
problem this commit deletes).

SEMANTICS CORRECTED. The WHOOP 1 Hz accel field is a fused GRAVITY vector, not
acceleration: across 269,486 real samples ||a|| is p50 1.027 g with 0.030%
above 1.3 g, and during the single most vigorous minute of a day it measured
1.033 g +- 0.006 with 0 of 420 samples above 1.2 g. So `dynAmp` measures how
fast the wrist RE-ORIENTS, and ENMO/MAD over this substrate reduce to
~(1.03 - gRef): a pure calibration artifact with zero signal. That is the true
root cause of the 42,155-steps-at-gRef-0.97 / 0-at-1.02 collapse, and it is
now documented on the type instead of a comment claiming gravity removal
leaves "motion".

388 tests pass.

…lamp

CodeRabbit findings, all three verified against the code before fixing.

  * The frozen-floor test PASSED VACUOUSLY. `day(0.05*k, 0.4*k, 200, 60)` held
    the active block at a fixed 60 minutes for every k, so the frozen count was
    60 on every iteration and `expect(f, greaterThanOrEqualTo(lastFrozen))`
    compared a constant to itself. It proved nothing while claiming to prove
    the central result of the PR.

    Rebuilt on a GRADED distribution, which is what real activity looks like:
    activity does not scale a fixed-size block, it lifts more of a continuous
    distribution above the threshold (the real-data proof moved 23 -> 254).
    Frozen counts now genuinely vary (125 -> 250) and the assertion is STRICT
    monotonic increase plus a 1.5x margin, so a regression to constant output
    fails loudly.

  * Stale pass-1 gate comment still described "movement inside the ambulatory
    band" and "(when HR is available) HR lifted off rest" above a loop that
    does neither. Same class of known-false comment this PR exists to remove.

  * `clamp(0.30 * clamp(...), 0.1, 0.45)` — the inner term caps at 0.30, so the
    0.45 outer bound was unreachable and implied a confidence this metric never
    claims. Aligned to the reachable value.

Also corrected the FILE HEADER, which review did not flag but was describing
the removed design: it still promised an HR corroboration gate, a "reported
step band", and a cadence that "NARROWS" that band. Now describes the frozen
floor, the deleted gates, and why cadence is measured but deliberately not
consumed by the 1 Hz path.

388 tests pass.