[EPIC] Verification goals and proof-run lifecycle

[Joncallim]

Execution mode: tracking
Parent Epic: #184
VNext programme: #333

Issue Type

Epic - verification-goal capability tracking. Do not dispatch this Epic directly to an implementation agent.

Context

#187 originally combined repository-backed goal definitions, on-demand execution and project-specific scheduling. The first three foundation slices are already merged on current main:

• PR feat: add verification goal registry foundation (#187) #328 — verification-goal registry foundation;
• PR feat: add authenticated verification goal imports (#187) #329 — authenticated import API;
• PR feat: add authoritative verification goal registry revisions #330 — authoritative registry revisions.

The old delivery branch/PR #331 then accumulated executable runner, database policy and scheduling work before the VNext Mission/Execution/Trigger architecture was frozen. That monolithic pre-VNext branch is now superseded as an implementation vehicle.

The remaining capability is intentionally split so verification goals do not own a second execution or scheduler runtime:

• [FEATURE] Complete verification-goal on-demand proof execution through VNext #355 — bounded on-demand proof execution through VNext Execution/Operation/Grant/confinement;
• [FEATURE] Schedule verification-goal proof runs through generic Triggers #356 — recurring proof delivery through [FEATURE] VNext Phase 7 — Trigger/Event runtime with dedupe, causality, and zero-token idle #341 generic Triggers.

Historical PR #331 remains evidence/reference for reviewed edge cases only; it must not be merged/cherry-picked wholesale.

Desired Outcome

Forge has versioned “what still works” goal definitions whose current revisions can be executed safely on demand and on a controlled cadence, with canonical outcomes/reliability evidence, last-known-green/first-observed-failure projections, and no speculative automatic repair.

Goal definitions own proof semantics. Generic VNext services own Execution, side-effect confinement/recovery, Missions and schedule/event delivery.

Tasks

• Registry/schema foundation — PR feat: add verification goal registry foundation (#187) #328.
• Authenticated import/read surface — PR feat: add authenticated verification goal imports (#187) #329.
• Authoritative immutable/current revision semantics — PR feat: add authoritative verification goal registry revisions #330.
• [FEATURE] Complete verification-goal on-demand proof execution through VNext #355 — complete on-demand proof execution through current VNext contracts.
• [FEATURE] Schedule verification-goal proof runs through generic Triggers #356 — schedule proof runs through generic Triggers after [FEATURE] VNext Phase 7 — Trigger/Event runtime with dedupe, causality, and zero-token idle #341.
• Confirm [FEATURE] Normalize execution outcomes and stop reasons #185 canonical outcomes and [FEATURE] Add capability reliability ledger #186 comparable reliability evidence are emitted by proof runs.
• Confirm [FEATURE] Add reliability, autonomy, and regression reporting #191 reporting consumes goal/run evidence without creating a new source of truth.
• Retire/close stale PR feat: implement project verification goals and proof runs (#187) #331 and all obsolete [EPIC] Verification goals and proof-run lifecycle #187 implementation branches after preservation of relevant requirements/tests.

Acceptance Criteria

• Goal registry/revision foundations on main remain authoritative and are not replaced.
• On-demand proof runs execute through [FEATURE] Complete verification-goal on-demand proof execution through VNext #355 VNext Execution/Operation/Grant/confinement semantics.
• Recurring proof runs execute through [FEATURE] Schedule verification-goal proof runs through generic Triggers #356/[FEATURE] VNext Phase 7 — Trigger/Event runtime with dedupe, causality, and zero-token idle #341 Trigger occurrences, with no verification-specific scheduler truth.
• Deterministic proof execution does not require an LLM.
• Goal/run history links to [FEATURE] Normalize execution outcomes and stop reasons #185 canonical outcomes and comparable [FEATURE] Add capability reliability ledger #186 evidence without forging missing/unknown evidence into success.
• Last-known-green and first-observed-failing state are rebuildable projections over immutable proof history.
• Disabled/archived/revised goals cannot be executed by stale requests/occurrences.
• Failed proof does not directly launch an unreviewed speculative code repair.
• PR feat: implement project verification goals and proof runs (#187) #331 is closed/superseded and no stale branch is treated as the current implementation base.

Out of Scope

• Owning Mission/Execution state machines.
• Owning schedule/cron infrastructure.
• Independent model verification ([FEATURE] Add independent Verification Workforce execution #188).
• Earned-autonomy policy ([FEATURE] Add evidence-based earned autonomy policy engine #189).
• Reporting/dashboard ([FEATURE] Add reliability, autonomy, and regression reporting #191).

Technical Notes

The reusable design lessons from PR #331 are bounded command/Resource policy, restart/idempotency tests, output/redaction limits, schedule dedupe test cases and run-history semantics. Reimplement these only where they still fit current VNext contracts; do not preserve old module/table structure merely because it exists on the branch.

[Joncallim]

VNext sequencing note (2026-09-03): keep the landed verification-goal definition/registry work as a trust foundation. Do not introduce a standalone project-only scheduler as the remaining scheduled-run implementation. First align the definition/run contract with #334; scheduled recurrence should ultimately consume the canonical Trigger/Event substrate in #341, and goal outcomes/evidence should continue to reuse #185/#186. Manual/on-demand goal execution can proceed earlier if it does not create parallel Mission/Trigger/orchestration truth.