DirectSound live rate-clock integration — 2026-09-22

The preceding WAT frequency-state fix exposed a separate host accounting bug: the cursor was elapsed * currentFrequency, so SetFrequency retrospectively changed how much audio had already played. A fake AudioContext regression doubled frequency without advancing time and observed the cursor jump from 8819 to 17639 bytes.

lib/host-audio.js now integrates piecewise-constant playback rates. Each changed clock retains its origin and rate boundaries with accumulated media time. Guest clocks retain millisecond units; audio clocks use seconds. This avoids rounding exact guest ticks through an early seconds conversion.

The shared calculation feeds snapshot cursors, queued-ring cursor estimates, stream cursors, one-shot guest-clock completion and the future splice offset. The AudioWorklet's published cursor remains authoritative when available. Ordinary fixed-rate voices take the arithmetic path without creating history. New Play clears history; changed splice/restart origins invalidate their old clock, and stream reset clears it explicitly.

Device-latency and deadline-lag queries may refer to samples before the most recent rate change, so simply rebasing a single start timestamp is insufficient. History retains the current delay window plus one older accumulated anchor. Changes at the same clock instant replace one boundary. Old boundaries are pruned on the next rate change rather than accumulating for the lifetime of a looping sound.

Validation

test/test-directsound-play-cursor.js failed before the change at the immediate 8819 -> 17639 cursor jump. It now covers immediate continuity, two rate changes, zero/100ms output latency, queries in an older still-in-flight segment, headless one-shot completion, new Play, stream reset, and 1000 same-tick plus 1000 advancing-tick edits with bounded history. Existing latency and WAT write-cursor checks remain intact.

Related checks: ring-refresh-seam (10), ring-queue, ring-wrap, AudioWorklet DSP (30), worklet routing (36), and loop-refresh pass. The loop-refresh test has concurrent worktree edits owned by another agent; it was run but not modified or included in this commit. waveOut audio/completion/reset and suspend-backlog (13 checks) also pass, along with test-tier membership and whitespace validation.

Limits / next work

• No real-device audio recording, native Win98 capture or performance number is claimed. This test verifies host clock accounting, not sample-perfect audible conformance across every backend.
• Existing queued source scheduling and waveOut completion deadlines were not redesigned. Changing frequency must eventually be checked against the actual samples and all pending sources, not just reported cursor position.
• Frozen-recording boundary retiming is now implemented separately in directsound-recording-rate-review.md. Reconstruction-level sub-frame timestamp precision remains unverified.
• Dynamic increases of device latency after history pruning and backwards guest-clock jumps are not certified by this test.
• Correct centibel pan attenuation, primary device-wide mixer controls and shared PCM ownership for duplicate buffers remain open.

The broader fable-review.md goal is still open; this addresses another stateful-looking implementation that returned plausible but wrong results.