Gate product delivery on the actual user workflow, not component test counts

[infotradescout]

Failure being corrected

An automatic companion was delivered as competing applications with manual-first surfaces, while component test counts and screenshots were presented as product-quality evidence. Existing SI doctrine prohibited this, but its application and product-delivery proof were not established. This repair does not claim that source publication changes every installed client.

Changes

• Add delivery_acceptance.py: current-intent/source/owner binding, exactly one intended runtime target, verified installation evidence, automatic workflow with zero manual-entry steps, scoped workflow/visual/restart/safety/launch evidence, explicit unresolved-rejection blockers, artifact integrity and safe evidence paths.
• Integrate it into the existing quality_gate.py --delivery path. The original five source checks remain separate and unchanged in purpose. Source validation explicitly reports product delivery as not evaluated.
• Link the product acceptance protocol from the existing evidence-and-completion authority.
• Add a bounded discovery-admission helper for known paths, pending searches and repeated no-progress searches. This helper requires a caller-persisted ledger; it does not intercept arbitrary external tool calls.
• Add 39 synthetic validator/CLI regression tests. The strongest verdict is ready_for_review, never user_accepted.

Verification executed

python -B -m unittest discover -s skills/selective-intelligence/tests -p test_delivery_acceptance.py -v: 39 passed, zero failures, in an isolated partial source workspace. Exact tested module blobs match this branch:

• delivery_acceptance.py: 80dba8024287c5ad3e79e16fe6c4cb8f27cd8aaa
• quality_gate.py: b2ddbcfb171a7b05dac70fd6cc9adaa23c7aab41
• test_delivery_acceptance.py: 6db87c87eff874a38a63f23c9c45c6a70699787a

A current companion repair packet was evaluated through the actual quality CLI and correctly returned needs_repair despite passing source/API and synthetic UI checks. Native installation and gameplay proof were not fabricated.

Not yet verified / merge blockers

• Full five-stage repository quality gate against this exact complete tree.
• Adapter/distribution regeneration and associated integrity checks.
• Required invocation by installed clients before their product-delivery state transitions; no system-wide interception is claimed.
• Native Windows and fresh-context installed-client forward tests.

The connected PC became unreachable during the repair. No installed SI runtime or game save was changed. Keep this PR draft and unmerged until the remaining checks are completed. Based on the exact PR #54 candidate 46629c3880151e5bcef090a5a85bbb7f8eeeaff4; it does not replace or merge that existing runtime repair.

[infotradescout]

Hosted verification checkpoint: Quality run 35252937420 is marked failed. Its Deterministic checks job 105309443344 returned an empty steps list; fetching its log returned 404 / BlobNotFound. This provides no executed full-suite evidence and does not establish a code failure cause. The 39 focused local tests remain the only executed SI validation in this repair. No check was disabled, no main branch was changed, and no installed-client enforcement is claimed. Full repository checks, distribution/adapter regeneration, and native forward validation remain merge blockers.

[infotradescout]

Installed-runtime checkpoint — September 17, 2026. This supersedes the earlier offline/partial-source checkpoint; PR remains draft and unmerged.

Published head: 07d00955001b7fcba8403407091676e4b585fb0c.

Completed:

• Registered the delivery evaluator and reference in the runtime/release manifests and the regression test in release files. Regenerated the canonical ChatGPT adapter: 61 files, one SKILL entrypoint.
• The original 59-file adapter budget remains enforced for all legacy files. Only the two explicitly reviewed delivery files are additionally admitted, and exact projection validation remains enabled.
• An attempted long core instruction failed the existing 1,100-word regression; the prose was reduced to a short mandatory product-handoff reference rather than increasing the limit.
• The unchanged five-stage source quality command passed on the exact published clean head. The adapter validation also passed. Full source quality is not product acceptance.
• Installed the 61-file canonical runtime allowlist on the authorized Windows PC at the existing .agents/skills/selective-intelligence location, preserved the complete previous installation outside skill discovery, and verified every installed file hash against the tested source. The existing AGENTS pointer still resolves the same canonical skill.
• A fresh native Python 3.11 process explicitly loaded the installed delivery evaluator and passed 33 selected evaluator regressions. The Windows symlink fixture and the source-only quality CLI test class were explicitly excluded from this focused installed-runtime check; this is not a full Windows suite claim.
• The app repair was separately installed as one canonical automatic-first companion with its old endpoint redirected, both journals preserved/reconciled, and native browser verification of map/reference/automatic capture.

Evidence retained on the owner machine:

• /tmp/si-delivery-published-quality-20260917.json in the isolated Ubuntu environment.
• .codex/project-execution/si-delivery-installed-20260917.json with exact source revision, file hashes, installation/backup identity and verified interpreter.
• .codex/project-execution/si-installed-delivery-tests-20260917.log with the actual installed-module path and 33 passing tests.

Remaining: complete native Windows acceptance including the recorded filesystem-capability cases; a fresh-agent end-to-end task showing enforced product handoff; broader client adoption. Runtime installation and a passing validator do not prove universal interception of arbitrary model/tool calls or subjective product quality. Infinity was not changed, and no game saves were written.