# Keep internal inference failures within their contract and budget

Status: contract cooldown, aggregate deadline and phase output allocations
verified; final installed full-source public-path acceptance passed.

The installed full-source Telegram harness, with a configured 32k context,
reached a real `virtual-memory:plan` request before its first work turn. Muse
returned HTTP 200, but its 1,000-token structured reply failed validation
(`transport_envelope`, `json_object`). A model-wide cooldown then suppressed
the different primary native-tool request without sending it. Intake failed
before any page was fetched. This is a production-path failure, not a mocked
provider response or a source-extraction failure.

Observed bundle SHA256:
`e1e3a9dfedc1a811497de0949c3d76ce3886f9cce6fd6227c903c9cf9758a716`.
Receipt: `/tmp/omnius-web-source-installed-public-live-20261013.json`.

1. Scope empty/invalid/expected-response validation cooldowns to the actual
   requested response contract and output allowance. A failed auxiliary JSON
   reply cannot suppress ordinary tool calling or a different response schema.
2. Preserve global availability, authentication and provider rate-limit
   cooldowns, including retry-after behavior. Do not change provider identity
   or bypass quota enforcement.
3. Verify failures remain cooled for equivalent contracts, while distinct
   work/review contracts and corrected output allowances can proceed.
4. Rebuild the installed candidate and repeat the real public full-source
   task. Keep the failed receipt, and record actual auxiliary terminal metadata
   without model text or credentials. Do not disable the failing phase merely
   to make the fixture pass.
5. Repair the confirmed underlying tiny-output defaults. Actual provider
   terminal receipts report `incomplete/max_output_tokens` with no visible
   output for both the 1,000-token VM planner and a 900-token artifact-search
   helper. Allocate adequate internal structured output, bounded by complete
   input, schemas, communication policy, context margin and explicit user caps.
   Preserve every input character; never fund output by clipping the request.
6. Fence the aggregate deadline after asynchronous preparation and route
   callbacks. An attempt timer clipped to the whole cascade's remaining time
   must end that cascade, including millisecond clock/timer skew; shorter
   per-attempt timeouts must still permit fallback.
7. Restore adequate primary-turn output before context admission. Reserve
   4,096 tokens, honoring an explicit lower user cap. Try the existing canonical
   owned-duplicate release before rejecting capacity; never clip protected
   source, request or authority. Remote eligibility must enforce the same
   host-owned reserve without sending that internal marker to providers or
   poisoning model-wide cooldowns when a route lacks capacity.
8. Keep completed extraction findings separate from its verbose checkpoint.
   Reopen all findings without rerunning completed workers. If the complete
   result still does not fit, return an explicit pending/capacity receipt.

An intermediate installed diagnostic run processed all 109,910 RFC source
characters in five valid workers, then exhausted its explicit 20-generation
test budget while small auxiliary requests fell through providers. This proves
whole-source processing, not accepted final delivery. A previous invalid worker
reply was not captured, so its exact rejecting field cannot honestly be inferred
from the later valid replies. It remains distinct from confirmed output
exhaustion and global cooldown poisoning.

The cooldown/deadline repair passes the original 94 remote-fallback tests,
11 contract tests and eight deterministic aggregate-deadline tests. Installed
live receipts show healthy main requests reaching Muse after auxiliary failures.
A broader legacy keyless-routing suite has 20 failures expecting older provider
priority; the same 20 failures reproduce with the prior committed fallback
source (`06443c0a`). Those fixtures are not counted as passing or changed here.

Subsequent live wire diagnostics distinguished a separate primary-turn limit:
662/882-token native tool requests failed after all source workers had already
completed. That was not the earlier 900-token artifact-query helper. Both
allocations are repaired without shortening their inputs. Final selected
runner/extraction/fallback/budget suites pass 264 tests with one existing
opt-in test skipped, including explicit user ceilings, provider route reserves,
unchanged inputs and completed-findings reopening.

The [final installed public acceptance](evidence/web-source-installed-public-live-20261013.json)
uses nine real model requests, all 109,910 RFC source characters and three
validated extraction workers. It produces a sourced final reply and settles
the native host goal after a scratch Telegram ACK. A separate
[32k run](evidence/web-source-installed-capacity-guard-20261013.json) proves that
insufficient capacity causes a typed before-transport pause with no tiny-output
request or automatic replay. It does not pass final-answer acceptance. No live
Telegram network delivery, local inference or fabricated provider response is
claimed by these receipts.

This repair supports the [complete-source work order](WO-web-source-full-coverage-extraction-20261013.md).
