Repository navigation
feat(OMN-19215): carry a lane-added backend's tier placement through the bifrost renderer - #3999
Conversation
…ugh to the rendered contract
The lane overlay binding accepts an optional placement {tier, fallback_for,
max_context_tokens} on an ADDED backend only, refuses one larger than the
backend's context_window, and the renderer passes it through unchanged. The
routing authority in omnimarket mirrors the backend into that tier after the
rungs it backs. An added backend with no placement renders exactly as before.
Onex-Lane: pc2-delegation-load
Onex-Session: 94b52a596fd34760ac199e09d09305b1
…ier behind the two .201 rungs Held: merge only after the omnimarket placement change is released into the dev lane image, since an older omnimarket refuses the placement key. Onex-Lane: pc2-delegation-load Onex-Session: 94b52a596fd34760ac199e09d09305b1
|
No OCC evidence companion was minted for this PR. product PR is a draft; companion suppressed (F-17) Matched in the state: To clear this: Mark the PR ready for review. The ready_for_review trigger re-fires the born path and the companion mints then. Reported by |
…h ProtocolConfigurationError Onex-Lane: pc2-delegation-load Onex-Session: 94b52a596fd34760ac199e09d09305b1
…s intact, and an oversized one is refused Onex-Lane: pc2-delegation-load Onex-Session: 94b52a596fd34760ac199e09d09305b1
#10983) * evidence: OCC companion pass 1 for OmniNode-ai/omnibase_infra#3999 * evidence: OCC companion self-bind for #10983 --------- Co-authored-by: node-occ-companion-effect <occ-companion-effect@omninode.ai>
✅ Hostile Reviewer — REVIEWEDCritical findings: 0
Semantics (OMN-17492 — the model finds, thread resolution gates)
Powered by omniintelligence.review_pairing.cli_review — multi-model adversarial review: qwen3-review (Qwen3.8-27B), gpt-oss-review (gpt-oss-120b) (OMN-8468/OMN-8524/OMN-17492) |
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=60&v=4){kind=small}
There was a problem hiding this comment.
Hostile Reviewer — adversarial findings (OMN-17492)
Models succeeded: glm-review
Models failed: codex
New finding threads: 0
Deduped (already posted on this PR): 0
Below quorum (one model only, reported not threaded): 7
Nit-level findings suppressed: 1
The model is the FINDER, never the gate: merge is gated only by the
deterministic Hostile Review Thread Gate, which blocks while
hostile-reviewer threads are unresolved. Resolve each thread after
addressing (or rejecting, with a reply) its finding.
Below quorum: 7 finding(s) raised by one model only (OMN-18479)
These are reported and NOT dropped, but they get no thread and do not block: a single model's finding no other model reproduced is not evidence enough to stop a merge. Read them; act on them if they are right.
- [MAJOR]
src/omnibase_infra/runtime/models/model_bifrost_lane_backend_binding.py (_validate_binding)(glm-review) — context_window comparison assumes non-optional field | The new validation compares self.placement.max_context_tokens > self.context_window. The diff does not show context_window's type; if it is Optio - [MINOR]
src/omnibase_infra/runtime/models/model_bifrost_lane_backend_placement.py(glm-review) — fallback_for may list the backend itself | Nothing in the new validation prevents placement.fallback_for from containing binding.backend_key. The routing authority mirrors the backend into the named t - [MINOR]
src/omnibase_infra/runtime/models/model_bifrost_lane_backend_placement.py(glm-review) — Duplicate entries in fallback_for are not rejected | fallback_for is only constrained to min_length=1. A tuple like ("local-coder", "local-coder") passes validation and would be passed through verbati - [MINOR]
tests/unit/runtime/test_bifrost_lane_overlay_placement_omn19215.py(glm-review) — No boundary test for max_context_tokens == context_window | The validation uses strict greater-than, so equality is legal. Both the unit and integration suites test only strictly-oversized values bein - [MINOR]
src/omnibase_infra/runtime/render_bifrost_delegation_contract.py(glm-review) — Cross-checks deferred to another package leave a render-time gap | The renderer passes tier and fallback_for through unvalidated, deferring rung-name checks to the omnimarket routing authority. Until - [MINOR]
tests/integration/runtime/test_dev_overlay_places_omnipc2_omn19215.py(glm-review) — Integration test couples to file paths and overlay contents | The integration test reads the committed dev overlay from a repo-relative path and hard-codes backend IDs, env var names, and the placed b - [MINOR]
src/omnibase_infra/runtime/models/model_bifrost_lane_backend_placement.py(glm-review) — max_context_tokens unconstrained against max_tokens | The validation checks placement.max_context_tokens against context_window only. A placement offering more context than the backend's declared max_
Lab readback, .201 dev lane, 2026-09-27 (OMN-19215 stack, with .202 down)Lane What ran. I hotpatched omnimarket#2913 at Condition. .202 is down. Its /data disk failed and llama-server cannot load its model (OMN-19863), so Result. I made six
A dead spread peer costs one refused connection and a same-tier sibling retry, not a failed delegation. So .202 does not need to come out of the overlay while its disk is out. Restore. The originals went back at 11:53:18Z. The deploy agent then recreated both containers from image Ordering hazard, measured. The lane's released omnimarket (0.4.245, the dev lane's current install) refuses the contract #4134 renders: Not proven here. These runs did not prove .202 answering (OMN-19215 AC2). That waits on OMN-19863. The receipts cannot show which host answered: both attempts carry the model-derived |
…es the current dev overlay omnibase_infra#4207 (OMN-17099) bound local-embedding in the dev overlay, and omnibase_infra#4005 (OMN-19251) removed local-ds-v4-flash from it. After dev was merged in, the minimal base contract this test writes still declared ds-v4-flash (a KeyError on the served-id lookup) and not local-embedding (the render refused it as an undeclared added backend). Declare the three base backends the overlay binds today, as omnibase_infra#4134 does, and type _dev_overlay's return so the file passes mypy --strict. The file is now byte-identical to #4134's. Onex-Lane: pc2-delegation-load Onex-Session: 94b52a596fd34760ac199e09d09305b1
|
OCC autobind rebound this PR's evidence-source stamp line to The receipt gate's own eligibility validator, run against the change-control tree the gate pins for Reported by |
There was a problem hiding this comment.
Hostile Reviewer — adversarial findings (OMN-17492)
Models succeeded: qwen3-review, gpt-oss-review
Models failed: none
New finding threads: 2
Deduped (already posted on this PR): 0
Below quorum (one model only, reported not threaded): 9
Nit-level findings suppressed: 2
The model is the FINDER, never the gate: merge is gated only by the
deterministic Hostile Review Thread Gate, which blocks while
hostile-reviewer threads are unresolved. Resolve each thread after
addressing (or rejecting, with a reply) its finding.
Below quorum: 9 finding(s) raised by one model only (OMN-18479)
These are reported and NOT dropped, but they get no thread and do not block: a single model's finding no other model reproduced is not evidence enough to stop a merge. Read them; act on them if they are right.
- [MAJOR]
src/omnibase_infra/runtime/models/model_bifrost_lane_backend_binding.py(qwen3-review) — Placement validation bypassed when context_window is None | The validator in _validate_binding checks self.placement.max_context_tokens > self.context_window. If context_window is None (the field is o - [MAJOR]
tests/unit/runtime/test_bifrost_lane_overlay_placement_omn19215.py(qwen3-review) — No test for placement with context_window omitted | The unit test file test_bifrost_lane_overlay_placement_omn19215.py covers oversized placement, missing rungs, empty tier, zero context, and unknown - [MAJOR]
src/omnibase_infra/runtime/models/model_bifrost_lane_backend_binding.py, src/omnibase_infra/runtime/render_bifrost_delegation_contract.py(gpt-oss-review) — Placement field introduces backward‑compatible contract change | ModelBifrostLaneBackendBinding now includes an optional placement attribute that is rendered into the delegation contract. Existing con - [MAJOR]
src/omnibase_infra/runtime/models/model_bifrost_lane_backend_binding.py, tests/integration/runtime/test_dev_overlay_places_omnipc2_omn19215.py(gpt-oss-review) — Inconsistent exception type for placement validation | ModelBifrostLaneBackendBinding._validate_binding raises ValueError when placement constraints are violated, while render_bifrost_delegation_contr - [MAJOR]
tests/unit/runtime/test_bifrost_lane_overlay_placement_omn19215.py, src/omnibase_infra/runtime/models/model_bifrost_lane_backend_binding.py(gpt-oss-review) — Unit test assumes placement validation is performed by Pydantic, but model raises ValueError | The unit test test_a_placement_larger_than_the_context_window_fails_naming_the_backend expects a Validati - [MAJOR]
src/omnibase_infra/runtime/models/model_bifrost_lane_backend_placement.py, src/omnibase_infra/runtime/models/model_bifrost_lane_backend_binding.py(gpt-oss-review) — Placement fallback_for references are not validated against existing backends | The new placement model only validates tier, fallback_for length, and max_context_tokens. It does not verify that each e - [MINOR]
src/omnibase_infra/runtime/models/model_bifrost_lane_backend_placement.py(qwen3-review) — Tier name in placement is not validated against known tiers | The placement.tier field is a free-form string with min_length=1. The docstring says the routing authority checks tier names against the l - [MINOR]
tests/integration/runtime/test_dev_overlay_places_omnipc2_omn19215.py(qwen3-review) — Integration test depends on the committed dev overlay file | test_dev_overlay_places_omnipc2_omn19215.py reads the actual docker/lane-overlays/dev.bifrost.yaml file. This couples the test to a mutable - [MINOR]
tests/integration/runtime/test_dev_overlay_places_omnipc2_omn19215.py(qwen3-review) — Negative-control test asserts target does not exist but renderer may write partial output | The test asserts not target.exists() after the renderer raises ProtocolConfigurationError. This is only corr
Findings demoted from threads (anchor rejected)
-
[MINOR] hostile-reviewer (qwen3-review)
fallback_for references not validated against the overlay's own backends | The placement's fallback_for tuple names backends that the routing authority will mirror into the tier. The renderer passes these names through without checking that they correspond to backends actually present in the rendered contract. The integration test asserts this post-hoc, but the renderer itself does not validate. A typo in a rung name will silently produce a rendered contract that the routing authority must reject, shifting
Resolve this thread when addressed — the
Hostile Review Thread Gateblocks while hostile-reviewer threads are unresolved (OMN-17492). -
[MINOR] hostile-reviewer (gpt-oss-review)
Use of assert for runtime validation | The function _added_backend_entry relies on assert statements to guarantee required fields (provider, tier, credential). In production, asserts may be disabled, causing silent acceptance of invalid data and later failures. | Evidence: src/omnibase_infra/runtime/render_bifrost_delegation_contract.py: lines containing "assert binding.provider is not None", "assert binding.tier is not None", "assert binding.credential is not None". | Fix: Replace asserts with explicit che
Resolve this thread when addressed — the
Hostile Review Thread Gateblocks while hostile-reviewer threads are unresolved (OMN-17492).
There was a problem hiding this comment.
Hostile Reviewer — adversarial findings (OMN-17492)
Models succeeded: qwen3-review, gpt-oss-review
Models failed: none
New finding threads: 2
Deduped (already posted on this PR): 0
Below quorum (one model only, reported not threaded): 7
Nit-level findings suppressed: 1
The model is the FINDER, never the gate: merge is gated only by the
deterministic Hostile Review Thread Gate, which blocks while
hostile-reviewer threads are unresolved. Resolve each thread after
addressing (or rejecting, with a reply) its finding.
Below quorum: 7 finding(s) raised by one model only (OMN-18479)
These are reported and NOT dropped, but they get no thread and do not block: a single model's finding no other model reproduced is not evidence enough to stop a merge. Read them; act on them if they are right.
- [MAJOR]
src/omnibase_infra/runtime/render_bifrost_delegation_contract.py(qwen3-review) — No validation that fallback_for rungs exist in the rendered contract | The placement model accepts arbitrary strings forfallback_for. The unit test `test_a_base_backend_cannot_be_given_a_placement_ - [MAJOR]
src/omnibase_infra/runtime/models/model_bifrost_lane_backend_placement.py(qwen3-review) — No validation that placement.tier matches the backend's own tier | A backend can declaretier: "local"while its placement saystier: "cheap_cloud". The model does not enforce that the placement t - [MINOR]
src/omnibase_infra/runtime/render_bifrost_delegation_contract.py(qwen3-review) — Placement data is passed through without sanitization or allowlist | Themodel_dump(mode="json")call on the placement object passes the tier and fallback_for strings directly into the rendered cont - [MINOR]
src/omnibase_infra/runtime/models/model_bifrost_lane_backend_placement.py(qwen3-review) — No test for placement with duplicate rung names in fallback_for | Thefallback_forfield is a tuple, so duplicates are possible:fallback_for: ("local-coder", "local-coder"). No test covers this c - [MINOR]
tests/integration/runtime/test_dev_overlay_places_omnipc2_omn19215.py(qwen3-review) — Integration test negative control mutates the real overlay in memory | The testtest_an_oversized_placement_in_the_dev_overlay_is_refusedloads the real dev overlay, mutates the placement in memory, - [MINOR]
src/omnibase_infra/runtime/render_bifrost_delegation_contract.py(qwen3-review) — No test for placement on a serving backend (serving=True) | The_added_backend_entryfunction setsendpoint_urltoNonewhenbinding.servingis True. A placement on a serving backend would be - [MINOR]
src/omnibase_infra/runtime/models/model_bifrost_lane_backend_binding.py(gpt-oss-review) — Inconsistent exception type for placement errors | When placement validation fails, the code raises a generic ValueError. The rendering pipeline and tests expect ProtocolConfigurationError (or Validat
Findings demoted from threads (anchor rejected)
-
[MAJOR] hostile-reviewer (qwen3-review)
Placement validation only fires for added backends, not base-declared ones | The _validate_binding method guards the placement check with
if self.placement is not None:but the inner logic only raises whennot declared. However, the context-window check (self.placement.max_context_tokens > self.context_window) is nested inside theif self.placement is not Noneblock but is NOT guarded by thedeclaredcheck. This means a base-declared backend that somehow carries a placement (which the first checkResolve this thread when addressed — the
Hostile Review Thread Gateblocks while hostile-reviewer threads are unresolved (OMN-17492). -
[MAJOR] hostile-reviewer (gpt-oss-review)
Placement validation method never executed | The ModelBifrostLaneBackendBinding class defines a private validate_binding method that checks placement constraints, but it is not registered with Pydantic's validation lifecycle (e.g., via @model_validator). Consequently, the constraints on placement (added‑backend requirement and context‑window limit) are never enforced at model instantiation, allowing invalid configurations to pass through to rendering. | Evidence: In src/omnibase_infra/runtime/models/model
Resolve this thread when addressed — the
Hostile Review Thread Gateblocks while hostile-reviewer threads are unresolved (OMN-17492).
…I Summary external-sweep false reds)
CI Summary read RED on two contexts that no later attempt of run 36350898404
ever re-posts: the bare unresolved 'Tests (Split ${{ matrix.split }}/...)'
(cancelled during attempt 1, 21:13:29Z, ci-watch class D2) and the bare
'deploy-gate' (skipped, from the pre-ready-flip push run at 21:09:56Z, also
D2). The compound 'deploy-gate / deploy-gate' is green and every real test
shard (Split 1/15..15/15) plus CI Tests Gate are green. Per pr-land/ci-watch
class D2, no rerun clears an orphaned bare-name context; only a fresh head
does.
Onex-Lane: pc2-delegation-load
Onex-Session: 94b52a596fd34760ac199e09d09305b1
|
OCC autobind rebound this PR's evidence-source stamp line to The receipt gate's own eligibility validator, run against the change-control tree the gate pins for Reported by |
There was a problem hiding this comment.
Hostile Reviewer — adversarial findings (OMN-17492)
Models succeeded: qwen3-review, gpt-oss-review
Models failed: none
New finding threads: 2
Deduped (already posted on this PR): 0
Below quorum (one model only, reported not threaded): 7
Nit-level findings suppressed: 1
The model is the FINDER, never the gate: merge is gated only by the
deterministic Hostile Review Thread Gate, which blocks while
hostile-reviewer threads are unresolved. Resolve each thread after
addressing (or rejecting, with a reply) its finding.
Below quorum: 7 finding(s) raised by one model only (OMN-18479)
These are reported and NOT dropped, but they get no thread and do not block: a single model's finding no other model reproduced is not evidence enough to stop a merge. Read them; act on them if they are right.
- [MAJOR]
src/omnibase_infra/runtime/render_bifrost_delegation_contract.py(qwen3-review) — No validation that fallback_for rungs reference backends in the same overlay | The placement's fallback_for field is a free-form tuple of strings. The renderer passes it through to the rendered contra - [MAJOR]
tests/integration/runtime/test_dev_overlay_places_omnipc2_omn19215.py(qwen3-review) — Integration test negative control mutates the real overlay file in memory but writes a poisoned copy | test_an_oversized_placement_in_the_dev_overlay_is_refused loads the real dev overlay, mutates the - [MAJOR]
src/omnibase_infra/runtime/models/model_bifrost_lane_backend_placement.py(gpt-oss-review) — fallback_for length validation ineffective | ModelBifrostLaneBackendPlacement defines fallback_for as a tuple with Field(min_length=1). Pydantic's min_length constraint applies to strings, not sequenc - [MINOR]
src/omnibase_infra/runtime/models/model_bifrost_lane_backend_placement.py(qwen3-review) — Placement tier and fallback_for are unvalidated free strings at the model level | ModelBifrostLaneBackendPlacement.tier is a str with min_length=1, and fallback_for is a tuple of strings with min_leng - [MINOR]
tests/unit/runtime/test_bifrost_lane_overlay_placement_omn19215.py(qwen3-review) — No test for placement on a backend that is not declared and not added (i.e., a base backend re-bound by the lane) | The unit test test_a_base_backend_cannot_be_given_a_placement_by_a_lane covers the c - [MINOR]
docker/lane-overlays/dev.bifrost.yaml(qwen3-review) — Dev overlay comment references 'the two .201 rungs it backs' but the fallback_for list has two entries | The YAML comment says 'placed in the local tier after the two .201 rungs it backs'. The fallbac - [MINOR]
src/omnibase_infra/runtime/render_bifrost_delegation_contract.py(gpt-oss-review) — Unnecessary type‑annotated variable in render function | render_bifrost_delegation_contract introduces a variable annotation 'entry: dict[object, object] = {...}' before populating the backend entry.
Findings demoted from threads (anchor rejected)
-
[MAJOR] hostile-reviewer (qwen3-review)
Placement validation only fires for declared backends, not for all backends | The context-window check in _validate_binding is gated behind 'if self.placement is not None' inside the 'if declared:' block. This means a base-declared backend that carries a placement will raise the 'not an added backend' error, but the context-window check is never reached for it. More critically, the check 'self.placement.max_context_tokens > self.context_window' is only evaluated when 'declared' is True. If a lane-added back
Resolve this thread when addressed — the
Hostile Review Thread Gateblocks while hostile-reviewer threads are unresolved (OMN-17492). -
[MAJOR] hostile-reviewer (gpt-oss-review)
Inconsistent exception type for placement validation errors | ModelBifrostLaneBackendBinding._validate_binding raises ValueError when a placement is supplied for a non‑added backend or when placement.max_context_tokens exceeds the backend's context_window. The rendering pipeline and tests expect ProtocolConfigurationError (or ValidationError) for configuration problems. Raising ValueError may propagate as an uncaught exception, breaking error‑handling contracts and leading to unexpected failure modes. | Evi
Resolve this thread when addressed — the
Hostile Review Thread Gateblocks while hostile-reviewer threads are unresolved (OMN-17492).
1 participant
Summary
OMN-19215, design (B), omnibase_infra half. The routing authority in omnimarket (OmniNode-ai/omnimarket#2803) mirrors a bifrost backend that declares a
placementinto its routing tier, after the rungs it backs. This PR carries that placement from the lane overlay into the rendered contract.ModelBifrostLaneBackendPlacement {tier, fallback_for, max_context_tokens}and an optionalplacementonModelBifrostLaneBackendBinding. Only a backend the lane ADDS may carry one (a base-declared rung is already in the ladder), and a placement offering more context than the backend'scontext_windowis refused naming the backend. Tier and rung names are checked by the routing authority against the ladder it loads._added_backend_entrywritesplacementwhen present and writes nothing new when absent, so every committed overlay without a placement renders byte-identical.local-omnipc2-chatin the local tier behindlocal-coderandlocal-heavy-reasoning, 32768 context (llama.cpp's per-slot window). It answers only when the .201 rung is unavailable or already tried.Sequencing
The dev-overlay commit must reach the lane only after omnimarket#2803 is released into the dev lane image: an omnimarket without that change refuses the
placementkey and the lane's delegation config would fail to load. Lab proof on the .201 dev lane (AC2: a test-scoped run whose rendered contract bindslocal-coderto a closed port answered by .202, plus an unmodified control answered by .201) runs after the M4 streak window closes. Nothing here is deployed before then.Tests
tests/unit/runtime/test_bifrost_lane_overlay_placement_omn19215.py(9): pass-through, no key when absent, context-window refusal naming the backend at the model and at the render (ProtocolConfigurationError), base-backend refusal, four malformed placements. Three were RED before the change (the binding refused the key as an extra field).tests/integration/runtime/test_dev_overlay_places_omnipc2_omn19215.py(2): the committed dev overlay renders the placement intact, every rung it names is bound in the same render, and a copy with an oversized placement is refused leaving no artifact.tests/unit/models/test_model_serialization_roundtrip.py: the new model has a factory and the binding factory now carries a placement.mypy --strictclean on the changed modules.Landing after M4
The orchestrator's M4 fence that kept this PR in draft has lifted (M4 was met 2026-09-23). This section records what landing requires.
Merge order (OMN-19215): 1)
omnimarket#2803merges todevfirst (the routing-authority half this PR's renderer output depends on). 2)release-on-merge.ymlauto-cuts the omnimarket release carrying that change. 3) this PR merges, consuming that published version in the dev-lane runtime image. 4) the AC2 lab pair (below) runs on the.201dev lane.Release cut required between omnimarket#2803 and this PR: the omnimarket auto-release from step 2 above. This PR's dev-overlay commit places
local-omnipc2-chatwith aplacementkey that only an omnimarket built from #2803 or later accepts; landing this PR's overlay commit against an older omnimarket would fail contract load on the lane.Lab proof needed on the
.201dev lane after M4 (not run today —.201is fenced): AC2 — a test-scoped run whose rendered contract bindslocal-coderto a closed port answered by.202, plus an unmodified control answered by.201, exercised through the deployed dev-laneonex-apicontainer.M4 probes to re-run after landing: C13, C14, C29 — this PR changes what the rendered delegation contract offers on the customer-local delegation path that those probes exercise.
.202 back in routing (2026-09-27)
The stop-using-.202 pause on this PR is released: operator RULING 2026-09-27T20:32:19Z lane=omnipc2-routing-restore (ledger:11849, "ok, looks fine I think. put it back into routing"), RELEASE rows ledger:11850, 11851 and 11855. .202 was rebuilt on its boot NVMe.
Serving profile reconciled, no change needed. .202 now serves
Qwen3.8-27B-MTP-Q3_K_M.ggufwith a q8_0 KV cache through a systemd drop-in (profile in the internal planning corpus). Read live 2026-09-27T20:3xZ:GET .202:8000/v1/modelsid and aliasQwen3.8-27B,meta.n_ctx32768, ftype Q3_K - Medium;/propstotal_slots 4;/slotsn_ctx 32768 on each of 4 slots; the unit still runs-c 32768 --parallel 4 --kv-unifiedexactly as before the rebuild. From insideomninode-runtime-effectson the .201 dev lane,GET .202:8000/v1/modelsreturns 200 withQwen3.8-27B, and a real chat completion to .202 returnsmodel: Qwen3.8-27B. This PR encodes the served id, the 32768 window and the four-slot limit, all unchanged; it encodes no model file, quant or KV type. Only the quant (IQ4_XS to Q3_K_M) and KV type (f16 to q8_0) changed, which no identity or health probe here checks.Lab proof
Hotpatch proof on the .201 dev lane, 2026-09-27 11:46-12:04Z (lab lease rows 2026-09-27T11:46:09Z and 12:04:16Z, lane delegation-fix-omnipc2-202-disk-and-routing-83, result PASS, restored=yes): omnibase_infra#4134 (whose commits include every commit of this PR) and omnimarket#2913 were patched into
omninode-runtimeandomninode-runtime-effects, the dev overlay rendered in-container, and sixonex delegateruns on kafka/dev/deployed-lane completed on local Qwen3.8-27B with 0 escalations. .202 was down then, so .202-hashed runs retried onto .201. That lane also checked offline that the released omnimarket loads this PR's render (it writes nomodekey). This PR's fallback-only render was not separately run on the lane.Lab pass still owed, with .202 live. Path: (1) the lab-pool proof on the head before it lands; (2) the compose-dev lab receipt keyed by the merged sha on the .201 dev lane; (3) AC2 through
onex-apion the .201 dev lane: a test-scoped rendered contract bindinglocal-coderto a closed port is answered by .202, and an unmodified control by .201.Order with omnibase_infra#4134. #4134 contains every commit of this PR plus spread mode. This PR can land first, because the released omnimarket already loads its render, and it puts .202 into the local tier as a fallback now; #4134 then lands after the omnimarket release carrying omnimarket#2913. If the drain prefers one PR, #4134 supersedes this one.
Branch updated from dev 2026-09-27; the dev-overlay placement test now matches the current dev overlay (
local-embeddingdeclared, the removed ds-v4-flash entry dropped) and passesmypy --strict(commit 45d719c). Focused suites green locally at that head: 103 passed; ruff, mypy and pre-commit clean on the changed file.Lab pool readback (OMN-18893)
Evidence-Ticket: OMN-19215
Evidence-Source: OCC#10983