Training storyboard proposal finalization loses committed state

[bokelley]

Problem

The current-source packaged storyboard path for media_buy_seller/proposal_finalize and media_buy_seller/proposal_finalize_asap_timing returns a committed proposal from the legacy get_products finalize step, but the following create_media_buy resolves the catalog draft and fails with PROPOSAL_NOT_COMMITTED.

This issue owns the proposal-state/session-handoff correctness bug. General storyboard process isolation, resource telemetry, memory limits, and teardown are tracked separately in #6778.

Reproduction

TENANT_PATH=sales PUBLIC_TEST_AGENT_TOKEN=storyboard-ci-token \
  npx tsx server/tests/manual/run-storyboards.ts \
  --filter media_buy_seller/proposal_finalize

Observed on 2026-08-22 with the current source overlay / packaged SDK path:

• proposal_finalize: 6 passed, 1 failed, 2 skipped
• proposal_finalize_asap_timing: 4 passed, 1 failed
• both failures: PROPOSAL_NOT_COMMITTED: Proposal "novamind_ai_audience" is a draft

The finalize response itself contains proposal_status: committed and an insertion order, so the failure is in the state/session handoff into create_media_buy, not the finalization response.

Likely boundary

Legacy get_products stores proposal state under getProductsSessionKeyFromArgs, while proposal-based create_media_buy can choose a principal-scoped sessionKeyFromArgs session and fall back to the global catalog draft. The fix must preserve cross-principal isolation while making the committed snapshot executable.

Acceptance

• Both proposal-finalization storyboards are clean in the current packaged runner.
• create_media_buy executes the committed snapshot returned by finalization.
• A cross-principal/account caller cannot execute another caller's proposal.
• Remove the temporary current-source quarantine and restore the associated sales passing-step floor when fixed.

Out of scope

• Child-process isolation, per-storyboard resource budgets, telemetry, and process-group teardown: Process-isolate storyboard execution and enforce per-storyboard resource budgets #6778.
• The separate packaged webhook and account-overlay failures: Packaged webhook_emission storyboard exhausts runner memory adcp-client#2653 and Packaged wholesale scope-isolation storyboard drops account scope adcp-client#2654.

[bokelley]

Triage

Classification: Bug — Conformance failure
Bucket(s): compliance-suite
Status: ready-to-implement

What the experts said:

• code-reviewer: Root cause confirmed: create_media_buy calls productDiscoverySessionKey with proposalRefinementScope absent, falling back to 'public_sandbox' account component, so the committed snapshot at open:p:<digest>:novamind_ai_audience is unreachable; OOM is a separate root cause — startSessionCleanup()'s 5-min interval never fires in a ~45s shard run, and lastGetProductsContext.products accumulates across all storyboard runs with no eviction.
• ad-tech-protocol-expert: Implementation bug only — spec already defines committed proposals as executable via create_media_buy; no normative surface touched, no changeset needed; companion Spec Backlog item: add MUST-language to create_media_buy task reference clarifying cross-task proposal visibility within a session (patch-eligible clarification, existing conformant implementations already satisfy it).

My take: Two distinct root causes, one PR. (1) Unify productDiscoverySessionKey's account component across get_products and create_media_buy dispatch paths so the committed snapshot is resolvable; preserve cross-principal guard (test two distinct principals against the same in-process store). (2) Call clearSessions() / reset InMemoryStateStore between storyboard runs and cap lastGetProductsContext.products to bound shard memory. P0 — media_buy_seller seller certification unconditionally fails accept_proposal until fixed.

Implementation scope:

• server/src/training-agent/task-handlers.ts lines ~6099, ~10459: propagate or unify proposalRefinementScope account component so create_media_buy resolves the committed snapshot under the correct key
• server/src/training-agent/state.ts: add inter-run store reset; optionally cap products retained in lastGetProductsContext
• server/tests/manual/run-storyboards.ts: call clearSessions() (or equivalent) between runs
• Required before merge: two-principal isolation test asserting neither can execute the other's committed proposal from the same in-process store
• No changeset (training-agent + runner only — not protocol package surface)
• Remove current-source quarantine and restore sales passing-step floor in same PR

---

Triaged by Claude Code. Session: https://claude.ai/code/session_01BiH6T2RjsJVcKpGdAQifKj

---

Generated by Claude Code

[bokelley]

Scope split: #6776 now owns only the committed-proposal session handoff and cross-principal correctness. Generic storyboard process isolation, RSS/time telemetry, resource budgets, and teardown moved to #6778. The temporary quarantine remains tied here until the behavioral failure is fixed; runner hardening must not hide that failure.