Fix criterion-bound Mission completion and durable successor planning

[pcvantol]

Problem and behavior

A canonical COMPLETE receipt previously proved every Mission criterion, so partial Action A could close the Mission without planning B. Forge now requires exact Architecture-approved evidence predicates, collects immutable delivery-bound JSON observations, and independently evaluates every criterion. Remaining scope and observed reasons reach the ordinary durable successor planner; one genuinely complete Action can still close directly.

The change preserves original observations and historical terminal results, distinguishes current-revision assertions from historical delivery, prevents repeated/no-progress work, and resumes the evidence-to-successor boundary without duplicate logical Actions or submissions. Schema 39 provides a reader fence while preserving all nonmetadata rows. The existing updater supports only the bounded 2.7.24→2.7.25 transition.

Validation

• Published 2.7.24 installed regression reproduced the evidence→evaluation→premature completion→missing planning chain.
• Full repository validation: 817 tests, one existing opt-in published-artifact test skipped; compilation, projection, version, JSON and whitespace checks passed.
• Eight composition scenarios use the actual installed factory/parser/evaluator/runner/storage with explicit external fixtures; exact candidate wheel qualification is recorded separately.
• Real Codex successor passed after a retained failed qualification and adapter correction: 11,942 input / 378 output tokens within existing bounds, one valid durable successor, zero live Host submissions. Usage enforcement is after generation, not a hard preflight ceiling.
• Real isolated EP owner HTTP handler/serializer and normal Forge adapter preserved authentication, candidate/delivery identities and artifact hashes; COMPLETE plus misleading PASS prose proved no criterion. Terminal transaction state was synthetic; no provider/control execution.
• Actual published old interpreter migration preserved all 45 nonmetadata table populations and rejected the resulting schema 39 when reopened with the old reader. Updater crash/recovery tests passed.
• Independent exact-candidate Quality/Security evidence will be recorded on this PR before protected merge.

Scope and delivery

Assignment L1-CRITERION-PLANNING-REPAIR-20260918. Version 2.7.25 selected once under the release policy. Release workflow qualifies exact wheel bytes before publication and repeats composition qualification against registry-downloaded bytes. This PR is implementation; finalization, normal release, safe updater activation and sanitized final handoff remain in this same assignment.

Repository assertions prove their approved structural properties only. Host command-control evidence is not available through current EP producer readback and remains explicitly unsupported; no EP source modification or invented host evidence. Mission 3, production reset and historical Mission reassessment remain excluded. #148 retains NIET_GEHAALD / NIET_GESTART.

[pcvantol]

Independent review and exact candidate qualification

Candidate: b368052ed59b18b19182ee724802f5d67d264340 (clean worktree).

• Independent Quality agent: PASS, no open findings;171 focused tests passed,1 skipped. All three prior findings were corrected and independently rechecked. Actual provider evidence matches this candidate's adapter digest.
• Separate independent Security agent: PASS, no open findings;50 focused tests passed on this SHA, with earlier contract/HTTP/migration/updater review also completed.
• Owning full validation:817 tests passed,1 existing opt-in published artifact test skipped; compile, version, projection, JSON and whitespace checks passed.
• Clean Git archive build of this candidate: wheel+sdist metadata/payload qualification PASS. Exact noneditable wheel installed away from source; all8 normal factory composition scenarios PASS across separate processes, including A→B, single-Action completion, misleading/missing/invalid evidence, no-progress, Action limit and later regression. Replay preserved state and effect counts.

Candidate-only artifact digests (not published release bytes):

• Wheel 3e288c76ec0b53a4aba795103eb72a53a7bb98667c7c9584ba87c02267762d45
• sdist 9cf5f07cf16464cf531cfe89cf6d5133f14562b94ff3c7b61d57acaf717d91d9

The reviewers did not author this implementation. This comment records their actual independent source assessments; it is not a fabricated GitHub review or product Mission assurance. Source review, candidate artifacts, protected finalization, published artifacts and installed activation remain distinct. The first failed real-provider qualification is retained; the corrected real invocation used11942input/378output within unchanged limits and produced one durable successor without a live Host call.