[mmastrac]

What

cargo test --release with the model pack present goes from roughly 30 minutes
to 3.5 minutes. The gates that were paying for that now run as a documented
real-model tier.

Why

The cost was the denoise step budget, not fixture size. A denoise step is about
1.3 s on an M3 Pro, so max_denoising_steps set the price of every model-gated
gate. The gates written first inherited SamplerConfig's 48-step default or
asked for 24, while the residue gates written later ask for 8. None of them
assert reply quality, and both arms of every comparison share one cfg, so the
step count cancels out of the assertion.

Changes

1. Step budgets to 8 across six gates, plus two token caps. Assertions untouched.
2. Nine gates that stay expensive at 8 steps, because they run two or more full
   generates or a lineage matrix of multi-thousand-token prefills, moved behind
   #[ignore = "real-model tier: ..."].
3. AGENTS.md documents the tier beside golden and smoketest.

Numbers

Per-test, serial, pack present, one process at a time.

test	before	after
per_block_ops_match_monolithic_generate	273.9s	100.7s	tier
ask_via_pipeline_matches_direct	254.7s	46.9s	tier
kv_lineage_unaligned_delta_offsets	234.8s	234.8s	tier
kv_lineage_paths_are_fingerprint_identical	218.5s	218.5s	tier
pipeline_rewind_kv_byte_consistency	201.6s	68.4s	tier
oplog_roundtrip_replays_bit_identically	132.6s	51.2s	tier
wrap_crossing_rewind_restores_state	84.2s	57.1s
truncate_after_uncommitted_canvas_write	83.6s	37.4s
tool_compact_m1_m2_and_overlong_smoke	72.0s	72.0s	tier
per_block_partial_commit_and_discard	69.3s	26.0s
forced_reroll_leaves_no_residue	61.9s	62.2s	tier
synthetic_kv_extension_byte_consistency_100k	39.0s	45.2s
tool_smoke_call_shape_and_argument_types	13.9s	13.9s	tier
begin_turn_after_abandoned_turn	12.1s	9.5s

gate	before	after
cargo test --release, parallel	not measured	211.3s, 794 passed
same, serial	~1830s reconstructed	244.3s
the tier	n/a	894.2s, 9 passed

The parallel before was never taken, because the first parallel run already had
the trims applied. The serial before is reconstructed from the per-test rows, so
it carries a tilde. Everything else is a stopwatch reading with no membudget
stalls in any reported run.

For the reviewer

• The tier command needs its --skip. Test filters match by substring, so
  kv_lineage_paths_are_fingerprint_identical on its own also selects the
  known-red q8 variant and the tier comes back failed. Caught by running it.
• The two kv_lineage matrices went to the tier rather than being trimmed. Their
  fork needs a wrapped ring, a misaligned resume and a multi-chunk delta together
  at specific offsets, so a smaller fixture would stop reproducing the bug
  without failing. That seemed worse than a slow test.
• begin_turn_after_abandoned_turn only prefills, so its step change is inert.
  The line is there to match its siblings and it buys nothing. Happy to drop it.
• synthetic_kv_extension_byte_consistency_100k is untouched. About 24 s of its
  39 s is FNV-1a over roughly 24 GB of live KV, four full layers at 100k tokens.
  A multi-accumulator hash would pay, but it changes every fingerprint written to
  op-log JSONL, which would make archived repro artifacts replay as divergent.
  That wants its own commit.
• Threading is unchanged. Before this split the suite was slower in parallel than
  serial, 987.7 s against 783.9 s, because the heavy gates contended for the GPU.
  With them in the tier, parallel leads again and the existing AGENTS.md note
  holds.
• Branched from aa6bff6d. main has since moved to f4b43d28.

🤖 Generated with Claude Code