Add PHC-base-clock refclock mode (phc.sync = false)

[jclark]

For #257.

With sync = false in the [phc] table, satpulsed leaves the PHC free-running and sends chrony SOCK samples whose tv is the PHC timestamp of each PPS edge, treated as TAI; the offset is the difference between GPS-derived TAI at the edge (with sawtooth correction) and that PHC reading. Chrony maps the PHC to the system clock on its side via its multi-clock support, so GPS-derived offsets are never re-expressed in the system-clock domain inside satpulse.

The branch carries the shared free-running machinery developed on phc-sample (pulse-edge admissibility, dual-edge polarity selection, the wallClock fit, PHC discontinuity detection, sawtooth correction), reconciled with current master, and reshapes emission to per-edge: the cross-sample regression, the sim rig, and everything else only the system-referenced flavor needs is deferred to #256, which stacks on this branch and ends up essentially where phc-sample was heading. With only one free-running flavor here, phc.sync = false alone selects the mode; the ntp.clock selector arrives with #256.

Observer.NTPSample does not fire in this mode: satpulse never computes a sys-vs-true offset when the PHC is free-running.

End-to-end validation against multi-clock chrony on real hardware is still pending (binary go/no-go per the plan in #257); this PR should not merge before that.

[jclark]

Review map: where the additive change is, relative to the old phc-sample branch.

Of the ~1040 non-test Go lines in this PR, roughly 600 are machinery carried from phc-sample unchanged or mechanically adapted, and the genuinely new logic is ~250 lines in five places:

1. phc_window.go, bottom half — labelEdge / labelEdges / Samples (~90 lines). The heart of the mode, replacing mapEdgesToUTC + fitAndEvaluate + TrueTimeOffset. Worth active checking:

   • Offset arithmetic: offset = (ref − edgePHC) − corrNs; the sawtooth sign follows from true_second = pulse_time + correction, so true time at the edge is ref − corr.
   • Precision: the potentially huge TAI−PHC distance is carried as an int64 Duration and converted to float64 seconds only at the end (the SOCK wire is a double anyway).
   • Consume semantics (new relative to 256's relabel-per-query): lastEmitted makes each edge emit at most once; a tolerance-rejected edge stays retryable until a later edge emits, then is skipped forever; errStale stops iteration mid-window; after warm-up the whole backlog emits in one burst.

2. generator.go — emit, MsgTAITime, leap piggyback (~40 lines). Emission triggers on both pulse and message arrival. One behavioral consequence to bless: in steady state the sample emits at the pulse, via the wallClock fit's forward extrapolation inside maxMsgGap (documented by TestGeneratorEmitsPerEdge). The leap flag attached to a sample is the last one delivered by the message stream.

3. timemsg.go additions (~80 lines). msgTAITime delivery: PrePulse skip, TAI-or-UTC-converted fallback (msgTAI via ls.UTCtoTime), millisecond rounding, monotonic dedupe, leap kind via StateAt. And GetTAIPulseCorrection: PrePulse-only, TAI-keyed walk-back, unrounded float64 via the new pulseOffsetNs split (whose rounded wrapper preserves the existing GetPulseCorrection behavior).

4. Dispatcher wiring (~50 lines). Per-mode sink routing; MsgTAITime is forwarded through the dispatcher so era transitions that swap d.generator don't leave the buffer pointing at a dead instance; Pause → NewInstance; deliberate absence of Observer.NTPSample in this mode.

5. Config/daemon (~20 lines). validatePHCSync (requires ntp.sock.path, excludes SHM) and gating ptp4l setup on phc.sync.

Everything else is deletion (regression layer, sim rig, extrapolation gate, leap-reset scaffolding — the TAI labelling is leap-immune), type-mechanical (the ntime.Time swap in wallclock.go; git diff --color-words origin/phc-sample -- time/internal/phcsample/wallclock.go collapses it to renames), or carried over verbatim (all the admissibility filtering, the wallClock gates, the config knobs). One semantic note on the "unchanged" wallClock: identical math, but its input changed domain — it now fits monotonic→TAI rather than →UTC.