feat(pld): hybrid partial-accept replay for SSM models (#134)

[st-adam]

Summary

• Fixes the token loss on hybrid SSM/ATT models when 0 < num_accept < K in PLD partial-reject path
• Adds Scheduler._replay_ssm_forward() to restore caches to N, replay accepted tokens, advance to N+K', emit K'+1 tokens instead of 1
• Default ON; opt-out: VMLX_DISABLE_PLD_REPLAY=1
• New /health field pld_ssm_replay.{enabled,attempts,emitted,failures}
• 6 unit tests in tests/test_pld_ssm_replay.py

Problem

PR #26 PLD on hybrid models (48 GatedDeltaNet + 16 full-attention): with K=2, a partial accept (num_accept=1) still emits only 1 correction token because SSM state cannot be trimmed — both caches must rewind to N.

Solution

After partial rejection, restore to N, replay drafts[:num_accept] forward through the full model. Both caches reach N+num_accept. Emit drafts[:num_accept] + [bonus_token] — same as the full-accept path, minus the extra K-K' tokens.

Expected gain

+5-10% on top of PR #26's +4-7% on hybrid models. Full PLD target for hybrid moves from +4-7% toward the +15-25% cited in #134.

Test plan

• pytest tests/test_pld_ssm_replay.py -v — 6 unit tests pass
• pytest tests/test_ssm_companion_cache.py -v — existing tests unaffected
• Live model: VMLX_DISABLE_PLD_REPLAY=1 vs unset — byte-equal at T=0, higher tok/s unset

Fixes #134

🤖 Generated with Claude Code

[st-adam]

Rebased onto current main (1.5.32 series, base commit 9cfbeb24) per @jjang-ai's note on #134 ("PRs in most recent stable versions"). No code changes — clean rebase over the 50 intervening commits.

Verified post-rebase:

• tests/test_pld_ssm_replay.py — 6/6 passing
• vmlx_engine.scheduler imports cleanly; Scheduler._replay_ssm_forward present
• Upstream's memory_limits refactor (v1.5.31) integrated cleanly into our scheduler.py touch points — no shadowing of get_metal_ws_guard_threshold / get_effective_metal_working_set_bytes

PR is CLEAN / MERGEABLE against the 1.5.32 base.