DeepSeek-V4.1-Flash arena preparation and Qwen-era deletions

[ShivanshuPurohit]

Workstream branch for the DeepSeek-V4.1-Flash B300 arena. Off-GPU preparation first; the GPU proof on the exact commissioned image follows on the same branch.

Commit 1: retire the recurrent prefix-cache audit and the Qwen-only wording

The prefix-cache state audit carried a recurrent (GDN) checkpoint branch and an int8 codec expression that no commissioned arena can reach: the Qwen arena seals disable_radix_cache=true, which the cache seam refuses for hybrid engines before any guard is built, and its deployed d9bd034 source predates the cache seam; GLM and the planned V4.1 arena have no recurrent pool. A hybrid pool that carries recurrent state is now refused at bind time instead of being served unchecked.

• Consumer: the tree_cache binding in cacheon/integrations/sglang_cache.py, which builds PrefixStateAudit on every hybrid engine.
• Casualty: the recurrent checkpoint audit and its mamba fixtures, the int8 codec expression, the Qwen recurrent-results section, and the Qwen-specific "disables radix caching" wording across the miner guide.
• Net lines: production +35/-99, tests +8/-159, docs +44/-78, scripts +3/-3.

Commit 2: draft the DeepSeek-V4.1-Flash arena

Read from the SGLang 0.5.21 source (tag e00930c5) and the published model and recipe facts; nothing here has run on a GPU yet.

• Node audit: DeepSeekV4TokenToKVPool stores whole 256-token pages, so the slot-addressed capture raised on its first call. cacheon/integrations/sglang_dsv4_state.py captures, restores and grades the window pages, a kv-source layer's compressed and FP4 index pages and its per-request pending-pair ring, decoded by the layout the pool declares; unrecognized layouts raise. Consumer: sglang_nodes._cache_rows. Supersedes nothing: GLM's DSA path is unchanged.
• Prefix cache: 0.5.21 hands a finished request through insert_req then on_release instead of cache_finished_req; the guard seals and checks both shapes, hashes the shorter DeepSeek-V4 index pages per full page, checks page-stored window pools by page, snapshots request-held compressor state across a handoff, and reads window eviction through the runtime's component API.
• Arena inputs examples/arena_inputs/dsv41flash/: Dockerfile on the v0.5.21 list digest, engine options (every key verified as a 0.5.21 ServerArgs field), 24 development requests in the GLM layout, README naming which node addresses execute (model.layers.* binds but never runs on this model).
• Dashboard: sources declare their URL slug in the operator config; the hardcoded arena slug table is deleted. The live sources.json needs the field at deploy.
• Net lines: production +238/-49, tests +256/-21, docs +23/-11, dashboard +14/-10, examples +144/-0.
• Not changed, recorded for later: GLM's DSA choice rule (count==3) is dead on 0.5.21, where the decoder layer returns a 2-tuple; it moves with GLM's own pin bump. The replay bridge needs no change: GLM's template also requires mapping tool arguments, so the sealed slice already carries them.

Commit 3: the owner rulings — TP2 with Engram host tables, batch-cell display retired, Qwen support retired

• Sealed engine environment: EngineSessionConfig.engine_env is a closed, reviewed name list (the two SGLang DeepSeek-V4.1 Engram host-table switches, which SGLang reads from the environment rather than ServerArgs), bound into the engine-config digest and forwarded by the session worker into build_session_environment, so every rank inherits it beneath the fence's own variables. cacheon check --engine-config honors the same engine_env object: exported before SGLang imports, never an Engine kwarg, recorded in the result. The kwargs allowlist gains the V4.1 options (ep_size, speculative_dspark_block_size, cuda_graph_max_bs_decode, swa_prefix_tails, prefill_decode_interval) and loses the Qwen GDN ones; the validator is generalized so oci_session_protocol.py stays at its ceiling. Consumer: the V4.1 kit's engine-config.json (TP2/EP2, memory fraction 0.85, per-rank host tables, about 190 GiB pinned host memory per engine) and the sealed arena definition. Casualty: the Qwen engine options and their fixtures.
• Batch-cell display retired (owner ruling Q-H): the dashboard, link previews and cacheon explain render replay results only — decode tok/s per user, first-token time, measured versus required gain, verdict. cacheon explain now reads the replay witness through the same reader as the dashboard, which also fixes the "not measured" line it printed for every V16/V17 product. Casualty: _phase_measurements, the B/C/B′ tok/s, prefill and TTFT/TPOT renderers, the winners' measured tok/s and prefill columns, the machine-noise headline (the replay grader's own noise bound decides).
• Qwen arena support retired (owner ruling Q-Q): a runtime that carries recurrent state raises in the node audit instead of being served unchecked; the GDN ReplaySSM rows, the Qwen input kit and the Qwen-specific doc passages are deleted. Kept: the H100 node-slot results page (tolerance evidence), the incident comments, and the H100/TP1 second test profile.
• Deploy notes: the arena prompt-authority engine_config is a closed field set, so any arena JSON (GLM's included, if it is ever re-commissioned on this code) carries "engine_env": {}; the dashboard sources.json carries slug (commit 2).
• Net lines: cacheon +112/-276, dashboard +34/-159, tests +179/-403, docs +35/-42, examples +24/-335, scripts +11/-10; every ratchet lowered to the measured counts (cacheon 90896, dashboard 4197, docs 12540, examples 368, tests 67240, dashboard/app.py 935) and the assertion floors of the three rewritten test files lowered with their production casualty.
• surface-change: modules (cacheon.chain.baseline_band exports replay_measurements; cacheon.integrations.sglang_dsv4_state registered with dsv4_state_rows).

Commits 4–9: H100 bring-up fixes (TP8/EP8 on 8×H100, 2026-10-08)

Integration on a box that is not the commissioned topology; nothing here is a B300 or qualification claim. Live controls through the public cacheon.cli check found four audit bugs in code that had never run on hardware, fixed at +19/-6 production lines:

• 12f1652: Engram commits its per-request token history outside the KV pools; the node audit now restores and grades it, so stock, the honest twin and the candidate start from the same n-gram.
• 3d5f76a: the compressor pending-pair allocation carries a spare ring, sentinel and alignment rows; both audits slice the request rows before reshaping.
• 76c1f0d: whole-node snapshots read the pool's active layer range instead of the model-wide mapping, which includes unowned draft entries.
• 3f41699: the prefix-cache read window starts at the page containing the live window, so a 128-token window inside 256-token pages is checked. The wrong-cache control passed before this fix.
• 6e3f630: the development requests submit the duplicate short batch after the first completes, so prefix reuse actually occurs (data only; concurrent duplicates produced zero hits).
• 0ab1ac1: GLM chain change — a completed remote NO_DECISION earns nothing and no longer blocks later PASSes from rewards or settlement. Deployed live on 2026-10-08; it rides to main with this PR.

Evidence, retained off-tree: honest MLP, attention+Engram, whole-model and prefix-cache checks PASS with eager audit and all-rank graph capture; the corrupted cache is rejected on all eight ranks; the wrong MLP fails with 80 violations; the honest cache re-passes after the boundary fix with 2,048 reused tokens in both phases. Engram host tables ran per rank (about 24 GiB per rank at TP8, fully huge-page backed).

Net lines for these commits: cacheon +39/-13, tests +132/-17, docs +12/-3, examples +18/-11, scripts +4/-4 (ratchet rows); the GLM commit is +76/-15 of that. surface-change: none.

CPU suite at fb4dcc8: 2839 passed, 11 skipped; 3 dashboard health tests fail on macOS for the missing /proc only. At 3f41699: 2836 passed, 11 skipped with the dashboard health tests deselected for the macOS /proc dependency; ruff, islands, line and assertion ratchets, the surface snapshot, the docs checker and mkdocs build --strict pass.

After these commits the production reference path ran on the same H100 engine: eight native agent turns from the sealed slice (1,060,616 prompt tokens, up to 178k each) generated to their caps under DSpark and graphs, the cache was flushed, and _reference_role_evidence scored all 10,172 tokens with finite logprobs; the generated tokens match the teacher's argmax at 98.3% of positions, the rest near-ties, so the token bridge and the teacher parser are aligned on V4.1. Unscored, retained off-tree, no production lines.

Remaining: the B300 day on the commissioned image — stock boot at TP2/EP2 with the host tables, paired lanes, stock-versus-stock replay noise, production controls, seal and calibration.

Commits 10–13: B300 commissioning of DeepSeek-V4.1-Flash (2026-10-09/10)

51a359d dispatcher memoization; 9dda14c Dockerfile layer order; 9cbce87 audit of registered controls; c54ccdd results page.

Dispatcher cost. The faithful identity control ran 4.4 to 8.0 percent slower than stock on the
3f41699b image. Timing the node seam inside the commissioned image put it at 85 µs per dispatched
call: a failed import of the removed legacy capture module on every call (28 µs), a receipt root
resolve on every completed call (35 µs) and descriptor plus variant match (12 µs). The fix (commit
51a359d) imports each capture authority once per process, memoizes the absolute receipt root, and
memoizes registry selection per graph mode, dtype, width and device: 9 to 12 µs per call.

Noise seal. The whole-model identity received PASS at +1.27 percent under the GLM-inherited
seal (1 percent per window, 0.4 percent per boot; pass line +0.74 percent). Both pairs now seal
window and boot noise at 0.02, the frozen policy's ceiling: pass line +2.2 percent at the final
look, +4.3 percent at the third.

Production-path proof on the commissioned image. First complete run of every stage: 75.3 min
(speed 55.0, eager audit 11.4 with 872 audited calls and 0 violations, pristine T 8.0). Futility
FAIL after one orientation: 28 min. Controls under the new seal, both pairs loaded concurrently:
identity FAIL on both pairs (window ratios 1.000/1.001/0.977/0.947 and 1.017/1.025/0.989/0.995,
pooled about +1.9 and −0.8 percent against the +2.2 line); candidate engine killed mid-window gives
infrastructure NO_DECISION three minutes later with the result retained; pod service restarted
mid-run: the restarted service's own recovery published the interrupted attempt as no_decision /
pod_service_restart without a hold, and the resubmitted wrong-MLP reached FAIL after four
windows at 1.007/1.015/1.012/1.020. Orientation-two skew toward the candidate while the neighbour
pair boots or tears down is recorded in the results page as the next noise item.

Audit of registered controls (9cbce87). Skip-MLP and wrong-MLP stop at the futility rule
because the sealed replay fixes token counts at temperature zero under DSPARK speculative decoding,
so a wrong answer can only be slower, and a sealed submission that fails speed is terminal: the
audit's rejection of wrong output had never executed on this arena. The sealed policy now carries
audit_control_delta_digests; a listed delta's speed non-PASS continues into the eager audit
(CALIBRATION_OBSERVATION), pristine T never runs after a failed speed, the verdict stays FAIL,
and every other delta remains terminal at speed. Builder, deployment check and intake accept the
disposition only for a listed delta; the key is serialized only when non-empty, so existing
policy digests are unchanged. Measured on the re-commissioned pairs: skip-MLP on both pairs
stopped speed at the first-orientation look (window ratios 1.024/1.051 and 1.058/0.954) and the eager
audit graded 34,880 calls with 34,800 violations (worst fraction 0.0000) -> FAIL / slot_audit_failed;
47 min from submission to verdict, 10.6 min of it the audit; pristine T did not run.

Operations. RAM checkpoint boot unit also reapplies persistence mode and the 1905 MHz clock lock
(systemctl start proof: Result=success, clocks 1905); retirement archives the whole pair tree
except replay slices and the VM pulls the archives; dashboard cut over to this source; CPU VM
intake, relay and supervisor units active on both pairs (activated 2026-10-10 01:29Z; the first two queued mainnet submissions transferred within a minute, one per pair).

Dockerfile. Dependency relocation moved above the source copy so a source-only rebuild reuses
the 20 GB layer. Before the reorder every build took 23 min; the first build after it 19 min
(export step 970 s); a source-only rebuild 37 s with the relocation layer cached.

Docs. docs/results/deepseek-v41-flash-b300.md (new), docs/validator-guide/arena-service.md
(reboot and retention procedure), qualification.md, fidelity.md, architecture/pipeline.md
and security/evidence.md (audit-control path), nav entry.

Net LOC for commits 10–13: cacheon +146/−69, tests +116/−40, docs +175/−6, examples
+6/−3, scripts +7/−7 (ratchet rows). Ceilings raised: cacheon +41 (dispatcher
memoization), examples +3 (Dockerfile comment), docs +15 (results page), root +1 (nav);
file ceilings qualification_runner.py 2595, test_qualification_runner.py 2014.

Commits 14–22: fee accounting and the deletion tranches (2026-10-11)

e0d4a03 fails candidate quality overlap under the sealed upper-bound policy. f4f9c48 and
9571c1a count evaluation fees per arena and total them across arenas on the dashboard; the
totals ride on /api/arenas and the existing arena fixture, and the separate revenue route from
the first cut is folded away (+304/−210).

Seven deletion tranches follow. Each was verified with the full CPU suite in a throwaway worktree
(editable finder stripped), ruff, islands, the line and assertion ratchets, the surface snapshot,
the docs checker and mkdocs build --strict. None changes a path the live intake process, the
standing supervisor, the pod service or the dashboard executes.

• a9cb2be: the retired bundles/dp_output_projection (its target is no longer registered, so
  TargetCatalog.require rejects it) and the untracked orphan worker-service test pinned in
  .gitignore. +1/−589.
• 70598f0: backfill_reservation_baseline_segments (no caller since 3685d41;
  commission_boundary superseded it) and the plain-store lease tests. +14/−344.
• 836f20f: test-only production symbols: release_manifest_compatibility_failure,
  settlement_blockers, the device-state capture verdicts, CohortPlan and _execution_order;
  the kernel-variant registry tests fold into test_capabilities. +45/−865.
  surface-change: modules cacheon.stack_plan drops CohortPlan; digest_domains
  cacheon/stack_plan.py keeps only cacheon.stack.marginal-arm-plan.
• de33c15: duplicated test fixtures fold into tests/support/evaluation.py and
  tests/support/b300.py; test_b300_arena_provider.py folds into
  test_b300_mainnet_worker.py; vacuous and duplicate tests in the runner, OCI backend, pod
  service and recovery tests are deleted. +596/−1443, no production lines.
• 93f7042: the stamp-1/2 intake migration and its schema-3 hold machinery
  (chain-archive-schema3-hold, the expiry and release guards). Every live intake database is
  schema 3 or later; the stamp-1/2 window closed 2026-07-13. +20/−384. surface-change: cli drops
  chain-archive-schema3-hold; modules cacheon.cli drops cmd_chain_archive_schema3_hold.
• f0b99e0: docs/architecture/overview.md and docs/engine/integration.md fold into the
  product-model and concepts pages. +37/−270.
• 31d333b: the validator loop's local qualification lane. chain-validate has only ever run
  --intake-only in production and now refuses any other mode; run_pass and run_validator
  lose the arena and mode parameters; ArenaServiceRegistry, AdmissionDecision,
  ArenaQueueSnapshot, admit_qualification, arena_queue_snapshot, qualification_cohort
  and --arena-id go with it. The intake path is unchanged apart from dedent;
  PassResult.decisions and settlements stay so the audit-log schema the dashboard reads is
  unchanged. +113/−639. surface-change: cli chain-validate drops --arena-id; modules
  cacheon.arena_service drops AdmissionDecision, ArenaQueueSnapshot,
  ArenaServiceRegistry and cacheon.chain.validator_loop drops QualificationWork;
  digest_domains cacheon/arena_service.py drops cacheon.arena.service-registry.

Deletion totals (a9cb2be..31d333b as one diff): production +56/−1,031, tests +627/−2,455,
docs +86/−398, bundles +0/−587, scripts and ratchet rows +35/−41. Every ratchet row and file
ceiling is lowered to the measured count. CPU suite at 31d333b: 2816 passed, 11 skipped, with the
three macOS /proc dashboard health tests deselected.

🤖 Generated with Claude Code