Local implementation follow-up for G38/PUL-5D98671B (PLAT-011/060 related) and G24/PUL-C004DFAF (PLAT-021/045/077 related). Not deployed. No live workflow configurations, decisions, or issue tracking rows were changed.

Review follow-up: published primary references retain their matching resolved provider, model, and options in runtime. The published reference stays on disk. Missing/mismatched resolution rejects the update before any write; regression tests cover usable primaries, rejection, and old-runtime isolation.

update_workflow_config(update_tier_fallbacks=...) now validates the entire patch against the fresh workflow.json, requires explicit routing mode and configured primary tiers, and writes the fallbacks durably before replacing the runtime tier configuration. Empty arrays intentionally clear fallbacks; invalid entries no longer silently clear or partially apply a list. Unrelated manifest fields are retained. Manifest persistence errors propagate to the caller rather than being logged behind a successful response.

Provider-profile mode is explicitly rejected, not silently converted. The operator's model-routing choice and any separate set_workflow_llm_config permission problem remain outside this fix. The server A report is not closed solely from this local change.

New dismiss_duplicate_human_input_request is registered in human_tools and the shared workshop/run human-tool surface. It requires workspace_path, input_id, keep_input_id, and reason. Both requests must be pending with identical source, question, context, options, free-text setting, run, evidence, and approval contract. The discarded ID must have no human_input_id reference in typed finding details or finding events. Malformed finding metadata fails closed.

The check, dismissal, and audit event are in one transaction. The retained request is untouched. No answer, consumption, or permission is fabricated. The duplicate_dismissed event retains the kept ID, reason, and session. This is not general agent authority to dismiss unique or linked user questions.

Focused local tests cover fallback persistence/clearing, profile rejection, invalid entries, missing primaries, denied writes, retained unrelated fields, registered duplicate-tool execution, linked/mismatched/non-pending rejection, retained pending state, absent fabricated answers, and audit history. Existing human-tool and human-input tests also pass.

The separate image-path report remains open for bridge-specific reproduction.