fix(headless): lock valid structured verifier failures as scored results

[me2seeks]

Summary

• treat valid structured verifier pass and failure outcomes as authoritative even when the agent cell exits unsuccessfully
• require the structured outcome, aggregate reward, and final verifier attempt to agree before overriding an infrastructure classification
• project stored structured failures as scored results so resumed runs do not resample them
• cover Harbor/Pier parity, infrastructure-only verifier outcomes, and WAL resume behavior

Closes#1585.

Verification

• npm run build:test
• npm --workspace @maka/headless test — 1,427 passed, 4 skipped because Harbor/Pier Python is unavailable, 0 failed
• npm run typecheck
• npm run lint
• npm run format:check
• git diff --check

Review focus

#1261 is a mechanical WAL-type extraction that also touches fixed-prompt-controller.ts. This PR keeps the scoring change and its regression coverage separate; I will rebase it onto the extracted type seam after #1261 lands.

[Astro-Han]

Approving current head 5765174 with a non-blocking P2 note.

Malformed or older verifier WAL data with a missing/non-array attempts field can throw a raw TypeError during projection and block resume. The shared verifier classification boundary should validate the runtime shape and map malformed data to the existing ungraded/infra path. The required test check is still pending.

[me2seeks]

Addressed the P2 in 3a8026f. The shared verifier classifier now treats runner and WAL payloads as untrusted runtime input, validates the harbor/verifier/attempt shapes before reading them, and returns the existing ungraded result for malformed data. The regression test covers both a missing attempts field and a non-array value; each resumes without resampling and remains scored=false / eligible=false.\n\nThe branch is rebased onto current main. Headless typecheck, lint, format, the focused 76-test controller suite, and the full Headless suite (1435 tests, 0 failures, 4 environment skips) pass locally.

[Astro-Han]

P2: taskCompletedEvent still treats any present harbor.verifier as scoring authority, even when structuredVerifierGrade() rejects it. On this head, { outcome: 'failed' } and non-array attempts both turn a max_tokens failure into task_completed with scored and eligible set to true. Sparse arrays also pass because Array#some skips holes. Please make the validated grade the only scoring input and cover malformed and sparse runner output.

[me2seeks]

Followed up in #1644. I removed the verifier-presence fallback, so runner data only affects scoring after structuredVerifierGrade() accepts the complete outcome/reward/attempt contract. The classifier now checks every array index explicitly, which rejects sparse attempts instead of letting Array#some skip holes.

The runner regressions cover missing and non-array attempts, a sparse array, and an internally inconsistent failed grade with a positive aggregate reward. Invalid verifier data remains ungraded, and a provider network failure still stays on the infrastructure path. The branch is rebased onto current main; build, lint, format, typecheck, and the full Headless suite pass locally.

[me2seeks]

Summary

• treat valid structured verifier pass and failure outcomes as authoritative even when the agent cell exits unsuccessfully
• require the structured outcome, aggregate reward, and final verifier attempt to agree before overriding an infrastructure classification
• project stored structured failures as scored results so resumed runs do not resample them
• cover Harbor/Pier parity, infrastructure-only verifier outcomes, and WAL resume behavior

Closes#1585.

Verification

• npm run build:test
• npm --workspace @maka/headless test — 1,427 passed, 4 skipped because Harbor/Pier Python is unavailable, 0 failed
• npm run typecheck
• npm run lint
• npm run format:check
• git diff --check

Review focus

#1261 is a mechanical WAL-type extraction that also touches fixed-prompt-controller.ts. This PR keeps the scoring change and its regression coverage separate; I will rebase it onto the extracted type seam after #1261 lands.

[Astro-Han]

Approving current head 5765174 with a non-blocking P2 note.

Malformed or older verifier WAL data with a missing/non-array attempts field can throw a raw TypeError during projection and block resume. The shared verifier classification boundary should validate the runtime shape and map malformed data to the existing ungraded/infra path. The required test check is still pending.

[me2seeks]

Addressed the P2 in 3a8026f. The shared verifier classifier now treats runner and WAL payloads as untrusted runtime input, validates the harbor/verifier/attempt shapes before reading them, and returns the existing ungraded result for malformed data. The regression test covers both a missing attempts field and a non-array value; each resumes without resampling and remains scored=false / eligible=false.\n\nThe branch is rebased onto current main. Headless typecheck, lint, format, the focused 76-test controller suite, and the full Headless suite (1435 tests, 0 failures, 4 environment skips) pass locally.

[Astro-Han]

P2: taskCompletedEvent still treats any present harbor.verifier as scoring authority, even when structuredVerifierGrade() rejects it. On this head, { outcome: 'failed' } and non-array attempts both turn a max_tokens failure into task_completed with scored and eligible set to true. Sparse arrays also pass because Array#some skips holes. Please make the validated grade the only scoring input and cover malformed and sparse runner output.

[me2seeks]

Followed up in #1644. I removed the verifier-presence fallback, so runner data only affects scoring after structuredVerifierGrade() accepts the complete outcome/reward/attempt contract. The classifier now checks every array index explicitly, which rejects sparse attempts instead of letting Array#some skip holes.

The runner regressions cover missing and non-array attempts, a sparse array, and an internally inconsistent failed grade with a positive aggregate reward. Invalid verifier data remains ungraded, and a provider network failure still stays on the infrastructure path. The branch is rebased onto current main; build, lint, format, typecheck, and the full Headless suite pass locally.

[me2seeks]

Summary

• treat valid structured verifier pass and failure outcomes as authoritative even when the agent cell exits unsuccessfully
• require the structured outcome, aggregate reward, and final verifier attempt to agree before overriding an infrastructure classification
• project stored structured failures as scored results so resumed runs do not resample them
• cover Harbor/Pier parity, infrastructure-only verifier outcomes, and WAL resume behavior

Closes#1585.

Verification

• npm run build:test
• npm --workspace @maka/headless test — 1,427 passed, 4 skipped because Harbor/Pier Python is unavailable, 0 failed
• npm run typecheck
• npm run lint
• npm run format:check
• git diff --check

Review focus

#1261 is a mechanical WAL-type extraction that also touches fixed-prompt-controller.ts. This PR keeps the scoring change and its regression coverage separate; I will rebase it onto the extracted type seam after #1261 lands.

[Astro-Han]

Approving current head 5765174 with a non-blocking P2 note.

Malformed or older verifier WAL data with a missing/non-array attempts field can throw a raw TypeError during projection and block resume. The shared verifier classification boundary should validate the runtime shape and map malformed data to the existing ungraded/infra path. The required test check is still pending.

[me2seeks]

Addressed the P2 in 3a8026f. The shared verifier classifier now treats runner and WAL payloads as untrusted runtime input, validates the harbor/verifier/attempt shapes before reading them, and returns the existing ungraded result for malformed data. The regression test covers both a missing attempts field and a non-array value; each resumes without resampling and remains scored=false / eligible=false.\n\nThe branch is rebased onto current main. Headless typecheck, lint, format, the focused 76-test controller suite, and the full Headless suite (1435 tests, 0 failures, 4 environment skips) pass locally.

[Astro-Han]

P2: taskCompletedEvent still treats any present harbor.verifier as scoring authority, even when structuredVerifierGrade() rejects it. On this head, { outcome: 'failed' } and non-array attempts both turn a max_tokens failure into task_completed with scored and eligible set to true. Sparse arrays also pass because Array#some skips holes. Please make the validated grade the only scoring input and cover malformed and sparse runner output.

[me2seeks]

Followed up in #1644. I removed the verifier-presence fallback, so runner data only affects scoring after structuredVerifierGrade() accepts the complete outcome/reward/attempt contract. The classifier now checks every array index explicitly, which rejects sparse attempts instead of letting Array#some skip holes.

The runner regressions cover missing and non-array attempts, a sparse array, and an internally inconsistent failed grade with a positive aggregate reward. Invalid verifier data remains ungraded, and a provider network failure still stays on the infrastructure path. The branch is rebased onto current main; build, lint, format, typecheck, and the full Headless suite pass locally.

[me2seeks]

Summary

• treat valid structured verifier pass and failure outcomes as authoritative even when the agent cell exits unsuccessfully
• require the structured outcome, aggregate reward, and final verifier attempt to agree before overriding an infrastructure classification
• project stored structured failures as scored results so resumed runs do not resample them
• cover Harbor/Pier parity, infrastructure-only verifier outcomes, and WAL resume behavior

Closes#1585.

Verification

• npm run build:test
• npm --workspace @maka/headless test — 1,427 passed, 4 skipped because Harbor/Pier Python is unavailable, 0 failed
• npm run typecheck
• npm run lint
• npm run format:check
• git diff --check

Review focus

#1261 is a mechanical WAL-type extraction that also touches fixed-prompt-controller.ts. This PR keeps the scoring change and its regression coverage separate; I will rebase it onto the extracted type seam after #1261 lands.

[Astro-Han]

Approving current head 5765174 with a non-blocking P2 note.

Malformed or older verifier WAL data with a missing/non-array attempts field can throw a raw TypeError during projection and block resume. The shared verifier classification boundary should validate the runtime shape and map malformed data to the existing ungraded/infra path. The required test check is still pending.

[me2seeks]

Addressed the P2 in 3a8026f. The shared verifier classifier now treats runner and WAL payloads as untrusted runtime input, validates the harbor/verifier/attempt shapes before reading them, and returns the existing ungraded result for malformed data. The regression test covers both a missing attempts field and a non-array value; each resumes without resampling and remains scored=false / eligible=false.\n\nThe branch is rebased onto current main. Headless typecheck, lint, format, the focused 76-test controller suite, and the full Headless suite (1435 tests, 0 failures, 4 environment skips) pass locally.

[Astro-Han]

P2: taskCompletedEvent still treats any present harbor.verifier as scoring authority, even when structuredVerifierGrade() rejects it. On this head, { outcome: 'failed' } and non-array attempts both turn a max_tokens failure into task_completed with scored and eligible set to true. Sparse arrays also pass because Array#some skips holes. Please make the validated grade the only scoring input and cover malformed and sparse runner output.

[me2seeks]

Followed up in #1644. I removed the verifier-presence fallback, so runner data only affects scoring after structuredVerifierGrade() accepts the complete outcome/reward/attempt contract. The classifier now checks every array index explicitly, which rejects sparse attempts instead of letting Array#some skip holes.

The runner regressions cover missing and non-array attempts, a sparse array, and an internally inconsistent failed grade with a positive aggregate reward. Invalid verifier data remains ungraded, and a provider network failure still stays on the infrastructure path. The branch is rebased onto current main; build, lint, format, typecheck, and the full Headless suite pass locally.

[me2seeks]

Summary

• treat valid structured verifier pass and failure outcomes as authoritative even when the agent cell exits unsuccessfully
• require the structured outcome, aggregate reward, and final verifier attempt to agree before overriding an infrastructure classification
• project stored structured failures as scored results so resumed runs do not resample them
• cover Harbor/Pier parity, infrastructure-only verifier outcomes, and WAL resume behavior

Closes#1585.

Verification

• npm run build:test
• npm --workspace @maka/headless test — 1,427 passed, 4 skipped because Harbor/Pier Python is unavailable, 0 failed
• npm run typecheck
• npm run lint
• npm run format:check
• git diff --check

Review focus

#1261 is a mechanical WAL-type extraction that also touches fixed-prompt-controller.ts. This PR keeps the scoring change and its regression coverage separate; I will rebase it onto the extracted type seam after #1261 lands.

[Astro-Han]

Approving current head 5765174 with a non-blocking P2 note.

Malformed or older verifier WAL data with a missing/non-array attempts field can throw a raw TypeError during projection and block resume. The shared verifier classification boundary should validate the runtime shape and map malformed data to the existing ungraded/infra path. The required test check is still pending.

[me2seeks]

Addressed the P2 in 3a8026f. The shared verifier classifier now treats runner and WAL payloads as untrusted runtime input, validates the harbor/verifier/attempt shapes before reading them, and returns the existing ungraded result for malformed data. The regression test covers both a missing attempts field and a non-array value; each resumes without resampling and remains scored=false / eligible=false.\n\nThe branch is rebased onto current main. Headless typecheck, lint, format, the focused 76-test controller suite, and the full Headless suite (1435 tests, 0 failures, 4 environment skips) pass locally.

[Astro-Han]

P2: taskCompletedEvent still treats any present harbor.verifier as scoring authority, even when structuredVerifierGrade() rejects it. On this head, { outcome: 'failed' } and non-array attempts both turn a max_tokens failure into task_completed with scored and eligible set to true. Sparse arrays also pass because Array#some skips holes. Please make the validated grade the only scoring input and cover malformed and sparse runner output.

[me2seeks]

Followed up in #1644. I removed the verifier-presence fallback, so runner data only affects scoring after structuredVerifierGrade() accepts the complete outcome/reward/attempt contract. The classifier now checks every array index explicitly, which rejects sparse attempts instead of letting Array#some skip holes.

The runner regressions cover missing and non-array attempts, a sparse array, and an internally inconsistent failed grade with a positive aggregate reward. Invalid verifier data remains ungraded, and a provider network failure still stays on the infrastructure path. The branch is rebased onto current main; build, lint, format, typecheck, and the full Headless suite pass locally.

[me2seeks]

Summary

• treat valid structured verifier pass and failure outcomes as authoritative even when the agent cell exits unsuccessfully
• require the structured outcome, aggregate reward, and final verifier attempt to agree before overriding an infrastructure classification
• project stored structured failures as scored results so resumed runs do not resample them
• cover Harbor/Pier parity, infrastructure-only verifier outcomes, and WAL resume behavior

Closes#1585.

Verification

• npm run build:test
• npm --workspace @maka/headless test — 1,427 passed, 4 skipped because Harbor/Pier Python is unavailable, 0 failed
• npm run typecheck
• npm run lint
• npm run format:check
• git diff --check

Review focus

#1261 is a mechanical WAL-type extraction that also touches fixed-prompt-controller.ts. This PR keeps the scoring change and its regression coverage separate; I will rebase it onto the extracted type seam after #1261 lands.

[Astro-Han]

Approving current head 5765174 with a non-blocking P2 note.

Malformed or older verifier WAL data with a missing/non-array attempts field can throw a raw TypeError during projection and block resume. The shared verifier classification boundary should validate the runtime shape and map malformed data to the existing ungraded/infra path. The required test check is still pending.

[me2seeks]

Addressed the P2 in 3a8026f. The shared verifier classifier now treats runner and WAL payloads as untrusted runtime input, validates the harbor/verifier/attempt shapes before reading them, and returns the existing ungraded result for malformed data. The regression test covers both a missing attempts field and a non-array value; each resumes without resampling and remains scored=false / eligible=false.\n\nThe branch is rebased onto current main. Headless typecheck, lint, format, the focused 76-test controller suite, and the full Headless suite (1435 tests, 0 failures, 4 environment skips) pass locally.

[Astro-Han]

P2: taskCompletedEvent still treats any present harbor.verifier as scoring authority, even when structuredVerifierGrade() rejects it. On this head, { outcome: 'failed' } and non-array attempts both turn a max_tokens failure into task_completed with scored and eligible set to true. Sparse arrays also pass because Array#some skips holes. Please make the validated grade the only scoring input and cover malformed and sparse runner output.

[me2seeks]

Followed up in #1644. I removed the verifier-presence fallback, so runner data only affects scoring after structuredVerifierGrade() accepts the complete outcome/reward/attempt contract. The classifier now checks every array index explicitly, which rejects sparse attempts instead of letting Array#some skip holes.

The runner regressions cover missing and non-array attempts, a sparse array, and an internally inconsistent failed grade with a positive aggregate reward. Invalid verifier data remains ungraded, and a provider network failure still stays on the infrastructure path. The branch is rebased onto current main; build, lint, format, typecheck, and the full Headless suite pass locally.

[me2seeks]

Summary

• treat valid structured verifier pass and failure outcomes as authoritative even when the agent cell exits unsuccessfully
• require the structured outcome, aggregate reward, and final verifier attempt to agree before overriding an infrastructure classification
• project stored structured failures as scored results so resumed runs do not resample them
• cover Harbor/Pier parity, infrastructure-only verifier outcomes, and WAL resume behavior

Closes#1585.

Verification

• npm run build:test
• npm --workspace @maka/headless test — 1,427 passed, 4 skipped because Harbor/Pier Python is unavailable, 0 failed
• npm run typecheck
• npm run lint
• npm run format:check
• git diff --check

Review focus

#1261 is a mechanical WAL-type extraction that also touches fixed-prompt-controller.ts. This PR keeps the scoring change and its regression coverage separate; I will rebase it onto the extracted type seam after #1261 lands.

[Astro-Han]

Approving current head 5765174 with a non-blocking P2 note.

Malformed or older verifier WAL data with a missing/non-array attempts field can throw a raw TypeError during projection and block resume. The shared verifier classification boundary should validate the runtime shape and map malformed data to the existing ungraded/infra path. The required test check is still pending.

[me2seeks]

Addressed the P2 in 3a8026f. The shared verifier classifier now treats runner and WAL payloads as untrusted runtime input, validates the harbor/verifier/attempt shapes before reading them, and returns the existing ungraded result for malformed data. The regression test covers both a missing attempts field and a non-array value; each resumes without resampling and remains scored=false / eligible=false.\n\nThe branch is rebased onto current main. Headless typecheck, lint, format, the focused 76-test controller suite, and the full Headless suite (1435 tests, 0 failures, 4 environment skips) pass locally.

[Astro-Han]

P2: taskCompletedEvent still treats any present harbor.verifier as scoring authority, even when structuredVerifierGrade() rejects it. On this head, { outcome: 'failed' } and non-array attempts both turn a max_tokens failure into task_completed with scored and eligible set to true. Sparse arrays also pass because Array#some skips holes. Please make the validated grade the only scoring input and cover malformed and sparse runner output.

[me2seeks]

Followed up in #1644. I removed the verifier-presence fallback, so runner data only affects scoring after structuredVerifierGrade() accepts the complete outcome/reward/attempt contract. The classifier now checks every array index explicitly, which rejects sparse attempts instead of letting Array#some skip holes.

The runner regressions cover missing and non-array attempts, a sparse array, and an internally inconsistent failed grade with a positive aggregate reward. Invalid verifier data remains ungraded, and a provider network failure still stays on the infrastructure path. The branch is rebased onto current main; build, lint, format, typecheck, and the full Headless suite pass locally.

[me2seeks]

Summary

• treat valid structured verifier pass and failure outcomes as authoritative even when the agent cell exits unsuccessfully
• require the structured outcome, aggregate reward, and final verifier attempt to agree before overriding an infrastructure classification
• project stored structured failures as scored results so resumed runs do not resample them
• cover Harbor/Pier parity, infrastructure-only verifier outcomes, and WAL resume behavior

Closes#1585.

Verification

• npm run build:test
• npm --workspace @maka/headless test — 1,427 passed, 4 skipped because Harbor/Pier Python is unavailable, 0 failed
• npm run typecheck
• npm run lint
• npm run format:check
• git diff --check

Review focus

#1261 is a mechanical WAL-type extraction that also touches fixed-prompt-controller.ts. This PR keeps the scoring change and its regression coverage separate; I will rebase it onto the extracted type seam after #1261 lands.

[Astro-Han]

Approving current head 5765174 with a non-blocking P2 note.

Malformed or older verifier WAL data with a missing/non-array attempts field can throw a raw TypeError during projection and block resume. The shared verifier classification boundary should validate the runtime shape and map malformed data to the existing ungraded/infra path. The required test check is still pending.

[me2seeks]

Addressed the P2 in 3a8026f. The shared verifier classifier now treats runner and WAL payloads as untrusted runtime input, validates the harbor/verifier/attempt shapes before reading them, and returns the existing ungraded result for malformed data. The regression test covers both a missing attempts field and a non-array value; each resumes without resampling and remains scored=false / eligible=false.\n\nThe branch is rebased onto current main. Headless typecheck, lint, format, the focused 76-test controller suite, and the full Headless suite (1435 tests, 0 failures, 4 environment skips) pass locally.

[Astro-Han]

P2: taskCompletedEvent still treats any present harbor.verifier as scoring authority, even when structuredVerifierGrade() rejects it. On this head, { outcome: 'failed' } and non-array attempts both turn a max_tokens failure into task_completed with scored and eligible set to true. Sparse arrays also pass because Array#some skips holes. Please make the validated grade the only scoring input and cover malformed and sparse runner output.

[me2seeks]

Followed up in #1644. I removed the verifier-presence fallback, so runner data only affects scoring after structuredVerifierGrade() accepts the complete outcome/reward/attempt contract. The classifier now checks every array index explicitly, which rejects sparse attempts instead of letting Array#some skip holes.

The runner regressions cover missing and non-array attempts, a sparse array, and an internally inconsistent failed grade with a positive aggregate reward. Invalid verifier data remains ungraded, and a provider network failure still stays on the infrastructure path. The branch is rebased onto current main; build, lint, format, typecheck, and the full Headless suite pass locally.