[AI] Add source-grounded publication summary candidates using shared runtime profiles #456

Outcome

Produce a useful, review-required publication synopsis from identified, approved textual sources, using a prepared local model or an explicitly permitted remote profile. Make the existing AI infrastructure usable in the comic workflow with a small reproducible demonstration.

Suite coordination: egohygiene/flow#80. Observability: Flow #14–#21. Domain behavior remains independently usable through Renderflow's CLI/library.

Reuse existing work

Renderflow #396, #309, #257 and #258 already own the model catalog, skill runtime, local/OpenAI-compatible adapters and caching; #274 owns existing timing/tracing. Generic summarization and skill.metadata.extract already exist. This issue adds a source-grounded consumer and operational profile integration, not another AI framework or generic summary transform. Related consumers: #349 repository briefings, #387 artifact DNA and #397 Sonic DNA; their scopes remain separate.

Checkpoints

1. Map the existing AI skill/catalog/config surfaces to the suite runtime profile. Reuse a working adapter; identify only the configuration/setup gaps.
2. Deliver one bounded text-to-synopsis candidate with structured evidence references and an explicit review/export step.
3. Demonstrate a synthetic comic script/transcript through a prepared local profile, record quality/resource results and integrate optional operational reporting. Remote-provider comparison is a separate explicitly authorized exercise.

Acceptance criteria

• Reuse existing skill/runtime/catalog/resolver, artifact store, hygiene, schemas, candidate approval and cache contracts. Consume the suite interoperability profile without a source dependency on Flow or sibling holons.
• Begin with approved authored script/transcript or textual companion inputs. Bound input/context/output and declare omissions; text-only input must not be represented as visual understanding of comic pages. Visual accessibility drafts require a separately declared compatible vision profile and review.
• Emit a separate versioned candidate with synopsis/summary, source artifact IDs/digests and bounded source-location references, template/schema/model/settings identity, uncertainty or omissions and review-required state. Validate reference membership and task-specific grounding separately from JSON shape; never claim schema validity proves factual correctness.
• The user can accept, revise or reject the exact candidate. Do not overwrite authored/canonical publication metadata or publish automatically.
• Support AI disabled, missing/unverified model, incompatible capability, local-only and explicitly selected remote profiles. Preparation/download remains separate from rendering. Missing local infrastructure never causes silent paid/cloud fallback.
• Reconcile documented environment/configuration/CLI mappings with the existing legacy transform and catalog/skill configuration surfaces. Resolve credentials through explicit handles; redact effective diagnostics. Keep existing deterministic forests and standalone operation intact.
• Apply bounded deadlines, attempts, token/output/cost policy and cancellation through the selected existing backend. Gateway behavior cannot expand permission, change models silently or defeat current retry/cache identity.
• Correlate run/stage/skill/model/profile, latency, attempts, cache decision, validation, candidate identity and reported usage with the existing observability track. Cost is reported, estimated with provenance, or unknown. Operational telemetry omits prompts, source text, private paths and secrets by default, and stays outside deterministic identities.
• Author fake-backend fixtures for no-AI, missing model, unsupported schema/capability, malformed output, invalid citations, source text resembling instructions, timeout/cancellation, changed model/profile/template cache identity and unavailable telemetry. Ordinary CI needs no paid endpoint.
• Publish a small redistribution-safe example, expected grounding/review criteria, measured local-runtime receipt when qualified, and known limits. Native/paid-model quality evidence remains separate from mock conformance.

Scope

No new router, autonomous multi-agent system, bundled unreviewed weights, model training, implicit networking, canonical metadata replacement or posting. This issue does not absorb #349/#397 or gate unrelated deterministic releases. Existing repository validation/qualification instructions remain in force; capture starts no execution.

Audit-to-roadmap reconciliation — 2026-10-21

Audit: https://github.com/egohygiene/renderflow/blob/audit/flow-suite-20261021T025401Z/audits/flow-suite-20261021T025401Z.md · Complete roadmap: #367

Source-grounded review-required summary consumer of existing AI runtime, with shared profiles; no new router or default remote fallback.

Bounded preparation: Design exact source slices/grounding/review schema and reuse existing runtime.

Full outcome remains gated (*): Shared runtime-profile/consumer compatibility, grounding/mock conformance and later measured local-runtime qualification; deterministic releases stay independent.

The maintainer's implementation-sprint qualification deferral remains in force. This audit ran no Renderflow tests, compiler/build/Cargo resolution, product lint/format/schema qualification, native/CLI/provider/visual execution, installers or new CI. Reading existing exact-source provider records is not a qualification rerun. Source implementation and authored tests are not passing-test, native-fidelity, release or downstream-consumer evidence.

This is an additive planning note. The full original issue, acceptance criteria, scope, parent references and historical checkpoints remain authoritative and unchanged. No acceptance box or issue is closed by this audit. Recommended ledger parent: #367.