Run goal scenarios through production task files and continuation admission

Draft
VA
10h ago
+1770-481

Problem

Goal tests previously performed continuation admission and cap writes themselves, so losing production automatic admission or its frames could pass. The private-goal catalog duplicated the runner. The complete family also exposed repeated durable-object history scans at the default 32-continuation boundary.

Design

Run all 45 declarative cases through LocalHarness request-file admission and production settle-to-continuation scheduling. Observe task-volume status snapshots/follows and scheduler frames; check model requests, exhausted scripts, exact final private/project files and cold-reopen replay without redispatch. Keep both default-limit boundaries and their five-second guards.

Fold #700's six private-goal histories into the shared driver. A completed project GOAL.md remains an independently asserted decoy while the private goal changes. Reject seeded project files without final expectations.

Own each agent fixture's workspace, durable state and settings under one temporary root. Resolve authority append and replay through the same immutable retry slot: only a retained rejection permits advancing; a committed slot terminates lookup. This fixes the production retry defect exposed by hermetic fixtures without scanning authority history.

Production performance prerequisite: #1015 removes the full object/part scan on every durable put. Its validation API accepts only the pending immutable body. The included signed merge preserves that independently reviewed production fix once.

Deletions

Delete the superseded goal catalogs, duplicate runners, obsolete output-volume override and test-owned admission/cap-writing logic. Excluding #1015: 80 files, +1362/-448, net +914 text lines, plus checked-in fuzz seeds. The authority fix adds 63 production lines; its remainder is verification. No added store, protocol, tool or model fake.

Decisions

Combine #600/#700 because they now use one driver and must pass one complete 45-case gate; close #700 as folded here. Keep #1015 separate as the production prerequisite. No closure-proof, task-segment decoder, request-history or growing-page experiments are included. Default-tools/fs main failures belong to main-green and are outside this focused gate. Goal-library tests remain in place.

Guarantee

  • Class: H/I-Is (growing history work and inadequate integration oracles), C/E (authority retry identity). Production admission, publication and persistence are exercised together through request files; the authority retry key is shared by append and replay, and a committed slot cannot be skipped.
  • Proof: complete retained goal family including both default-cap boundaries; unchanged-expectation control disabling production auto-admission; oracle controls for files, project files, frames, status, model context, request/task counts, script exhaustion, policy and output. Shrinking case-codec round trips, u32 continuation-boundary properties, parser/scenario fuzz targets and checked-in corpus cover the data path. Authority tests use the real crate-local Loom shim, two racing actors and preemption bound 2, with the old retry-key negative control; slot properties and native lost-ack settlement cover identity/replay. #1015 adds exact production N/2N body-read and allocation proofs against an independently initialized old scan, whose restored work fails the same predicate.
  • Entry points: LocalHarness request-file consumption, resume/settle automatic continuation and cold reopen; private and project goal files through the shared goal driver; ordinary/native authority append and replay; native lost-ack settlement. #1015 enumerates ordinary/native batches, mutation, multipart and recovery entry points for the shared persistence check.
  • Gaps: none.

Tests run

At signed 74da005e73e629ddc671edbe6620edb418cdf3a2, bb2 job1792317896451-1076648: all45 goal scenarios PASS3.882s (3.932s process), default-at PASS3.506s, default-before PASS2.905s, all6 goal properties/oracles PASS1.739s. Both reviewed default boundaries retain their five-second guards. Native compilation297.596s is excluded. This closes the workload failure in R600-010638.1 and R700-010643.1 through the complete family, rather than a private-only gate.

The extra agent-family gate failed10/16 in2.067s with workflow/workspace conflicts and indeterminate child operations, rather than deadlines. The guarantee matrix already identifies this production class as F8/F24.children CHILD-state, owned by subagents-c. The child/runtime production paths have no changes between #955 main and this head. The main-gate/e516e5986.log baseline independently records the same workspace-generation, indeterminate-child and uncertain-provider request-count failures (including wait-all, merge, message-while-running/paused and uncertain-provider). No agent expectation is weakened or removed. The continuation job1792318785410-1133719 passed the Harness/filesystem Windows GNU lib/tests check in346.545s. Its retained-binary dispatch trace failed7/16 agent cases in2.779s, showing Conflict("workspace generation changed") converted into indeterminate model commands, alongside rejected admitted calls/artifact publication. No full shared-suite PASS is claimed. Both failing runs and the independent baseline are retained; this is not a retry-until-green claim. See goal-complete-2eidmv5t/agent-fixture-family.log, agent-trace.log and Windows-GNU.log.

Unchanged authority proof reused from the reviewed source: real-shim Loom 3 passed in 0.0507s, slot properties 2 in 0.049s, authority regressions 8 in 0.071s, native original-lost-ack settlement 1 in 0.310s. No branch limit, budget or deadline was raised. #1015 at ca9b6e6dd: four new proofs 0.740s, quota regression 0.012s, replay property 0.157s; real HTTP/headless 33-request profile 4.590s before to 2.578s after, durable-put spans 990ms to 269ms. Objects Windows GNU compile passed 22.751s. The explicitly deferred macOS allocation-oracle live step is recorded in #1015.

Command: Q:/lanes/bs.ps1 python3 -u -c (Get-Content -Raw Q:/lanes/gc-e2e-goal-merged-gate.py). Native build: cargo test --locked -p acyclic-harness --test e2e --features acyclic-harness/filesystem-local,acyclic-harness/native-execution,acyclic-harness/mcp-http,acyclic-harness/grpc,acyclic-harness/terminal,acyclic-harness/search-http --no-run. Reruns execute the copied binary directly. Cross-check: cargo check --locked -p acyclic-harness -p acyclic-fs --lib --tests with that feature set and --target x86_64-pc-windows-gnu. Fuzz targets are Linux-only.