[ealtun21]

The long-full-eq@96000x8 scenario lost 768 frames once at a fixed frame after the soaks, then matched bit-exactly again, shifted.

Root cause: the recording Vec grew by doubling inside the WASAPI loopback capture callback. It outgrows its 120 MiB block at frame 3,932,160 (3840 * 2^10) and reallocates to 240 MiB; the callback stalled up to 28 ms (normally 1-3 ms), the engine's loopback buffer overran and dropped audio. Product (APO, daemon) was not involved.

Changes:

• native.rs: reserve and page-commit the whole recording before the stream starts; log the slowest capture callback and buffer growths.
• compare.rs: a run that diverges and is then re-aligned elsewhere reports 'slip of N frames (skipped) at frame F' instead of a generic mismatch (unit test added). The exact check is unchanged.
• windows runner: failing scenarios also save apo.log.
• spec section 16 documents the finding.

Evidence: old code on --scenario 'eq' reproduced the failure (slowest callback 28.7 ms, slip at frame 3931584); with the fix the same subset passes (slowest callback 12.5 us, 0 growths); full Windows tier 25/25 pass. One earlier full run with the fix had a single soak-flat dropout (callback 84 us, not the recorder) that did not recur.