Align D4 inline admission before freeze (#488)

dev-reports/issue-488/design.md

@@ -0,0 +1,72 @@
+# Issue #488 design
+
+Baseline: `85dd5fdd12b831deb49e7b3fe9e996c01f1fab67`, clean dedicated
+`feature/issue-488-cli-acceptance-d4-inline` worktree. No predecessors assigned.
+Read the complete parent-saved Issue body, worker skill and development guardrails.
+Parent `e99f1ebe` plus WIP and historical C5 `f0aceed8` are separate contexts;
+neither is an implementation source to copy or cherry-pick.
+Relevant committed ancestors inspected: `3c814d91` (#479 verifier formation),
+`d9ad6f5d` (#480 executed failure predicates), and `7125ce93` (#484 package-script
+formation). `HEAD` and the dedicated worktree's `origin/develop` both resolve to
+the full baseline SHA above; no parent mixed-worktree changes were assumed merged.
+
+## Existing coverage and reproduction plan
+
+The actual Runner calls Admission after normalization, augmentation and policy
+sanitization, before lint, and uses Admission::finish on every fallback return.
+#479 already rejects weak inline commands here using the final classifier; #484
+adds captured package-script obligations. Those changes are in this baseline.
+The three registration paths are register_step_plan_commands,
+verifier_obligations::register and complete_generated_verify_commands (handoff).
+They validate command policy but do not check immutable weak evidence.
+The baseline has no source_verifier_admission or C5 fixed_verifier_routing module.
+Existing profile and final-acceptance gates remain authoritative and unchanged.
+
+Before production edits, reproduce exact D4 command classification, Admission
+retry, rejected requirement-preserving rewrite and registration bypass with real
+functions. Preserve command bytes/hashes from the read-only historical input;
+use new reduced marker fixtures, not the disqualified historical argv recorder.
+
+## Smallest change
+
+Keep the existing broad formation rejection. Add a leaf registration guard for
+new final-success literal inline Node candidates only when trusted Config says
+Next.js/create, current-run provenance owns the generated unconfigured contract,
+and normalized trusted requirements make their source-independent weak evidence
+gate relevant. Use the final classifier and requirement evaluation, not a second
+weakness heuristic. Validate the entire candidate batch before persistence or
+producer-obligation updates. Existing configured/closed handling is unchanged.
+
+Extend preclosure projection with a closed whole-program grammar for literal
+file marker predicates: ordered literal-regex throw/assert or swallowed includes
+assert to direct includes assert. Preserve path, all ordered conditions, messages,
+success stdout and expected result; only literal regexes with escaped metacharacters
+qualify. No arbitrary JS equivalence. The existing model/host scope projection
+must still enforce outputs, responsible owners and execution boundaries. Record
+original command, normalized command, source step and checked replacement.
+Allow this projection only for owned eligible draft checks, never frozen checks.
+
+## Verification
+
+New corpus and focused Rust tests cover exact four commands on normal, four
+single-missing-marker and missing-file fixtures with real Node; final static
+classification and actual execution remain separate. Mock-provider tests capture
+actual request bodies and returned plans for repair, exhaustion, scope loss and
+fallback. Test both mixed candidate orders and all-good batches at all registration
+entries; authority/requirements controls; existing profile gates, independent
+business failure, compile/missing requirements and identity preservation.
+Run focused tests, related corpus, fmt check, clippy with warnings denied and full
+cargo test. Record failures honestly. No live model, browser, build-success claim,
+parent WIP edit, runtime namespace change, push or external lifecycle operation.
+
+## Baseline dynamic result (before production edits)
+
+`cargo test issue488_baseline --lib -- --nocapture` passed on the baseline with
+only corpus/test additions (1 test). Actual Admission returned Retry for O/V1/
+SWALLOW, Ready for V2; StepPlan registration accepted all four; final static
+acceptance rejected O/V1/SWALLOW and passed V2. The V2 rewrite after a weak first
+proposal was rejected by exact-command preservation. Thus normal weak refusal
+already exists, but registration defense and supported rewrite formation are
+unmet. No C5 routing transplant is needed. First test compilation failed because
+the test supplied typed deferred requirements to the string API; corrected the
+empty fixture argument before obtaining this result.

dev-reports/issue-488/implementation-summary.md

@@ -0,0 +1,47 @@
+# Issue #488 implementation
+
+At baseline `85dd5fdd12b831deb49e7b3fe9e996c01f1fab67`, #479 already returned
+D4 O/V1/SWALLOW to proposal repair. The unmet boundaries were direct registration
+and accepting a supported requirement-preserving rewrite. The pre-change dynamic
+reproduction is recorded in design.md; this is not the old C5 behavior replayed
+as if it were current develop.
+
+- Added a small registration guard at command, StepPlan/producer and empty
+  handoff registration. It validates the entire normalized/policy-checked batch
+  before contract persistence, producer changes or success events. The existing
+  final classifier is run without workspace files and its weak reason passes
+  through final acceptance's requirement normalization/pruning.
+- Scoped this guard to trusted Config Next.js/create, a same-run owned generated
+  unconfigured contract, new final-success inline literal candidates and a
+  gate-relevant weak reason. Existing broader formation rejection is retained.
+  Both profile checks use the existing `profile::is_nextjs_profile` boundary;
+  the final-acceptance test module uses the repository's cfg(test) convention.
+  Generality guardrail expectations, scanner and baselines are unchanged.
+- Added finite whole-program literal-marker formation: regex literal throw/assert
+  or swallowed includes predicates may become direct includes assertions only
+  with identical file, ordered conditions, diagnostic messages and success stdout.
+  Existing model/host scope projection retains other commands, outputs, responsible
+  owners, expected results and boundaries. Raw proposals cannot hide a shell
+  wrapper behind sanitizer normalization.
+- Reused the existing three planner proposals and fallback finish checks. The
+  new guard also checks final candidate returns. Added optional replacement
+  metadata to the existing event; no old event name/field/schema was removed.
+- Added exact-command/hash corpus, 24 reduced runtime cases, full final-executor
+  controls, captured mock requests/returns, registration all-or-nothing tests,
+  authority controls, synthetic independent business/compile failure boundaries
+  and frozen-byte identity rejection. Actual last-valid returns and clearing
+  are separate from direct finish tests.
+
+No final classifier promotion, requirement downgrade, added retry/Recovery
+budget, new fixed-verifier routing, parent WIP transplant, guardrail baseline
+change, live runtime namespace change or external action is part of this change.
+The baseline has no parent-WIP source_verifier_admission module; existing Next.js
+profile/build/behavior and identity checks remain separate. V2 passing a text
+check does not establish an application's business behavior.
+
+The evidence JSON files are curated test observations and captured request bodies,
+not historical or live run logs. Node v24.1.0 ran locally. `classifier_parser_argv`
+is the argument vector supplied to the direct Node control. Only status/stdout/
+stderr are compared with shell execution; shell-process argv is not observed.
+No live model API or browser experiment was performed. The new D4 measurements
+do not establish a successful application build or business workflow.