test(gateway): end-to-end accept trace over real SQLite rows and the real routes (composition)

[LamaSu]

Draft, test-only, stacked on #415. The composition path end to end over REAL SQLite rows, through the REAL HTTP routes. The product pack's return asks for "one end-to-end trace proving the lane's contribution against real state, not fixtures". Until now the traces held live rows as in-memory arrays.

What is real

• The store: the gateway's own store (initStore, getRepos, getStore), a real SQLite database migrated by the real createStore.
• Rows: capability and kernel rows are inserted, and later price-updated, through the real repositories.
• Re-read: productionAgentPlanDeps().revalidation runs unchanged, both for /api/agent-plans/validate and inside the accept seam. It is the real mapping from DB rows to live terms, over the real CSD registry (getCsdRegistry()).
• Routes and pipeline: the real agentPlanRoutes and reservationRoutes, the accept seam, the deal binding and presentPlan. Layer B and sealed come from the code; nothing is asserted by hand.
• Reservations: the durable BudgetReservationStore over the SAME SQLite handle, with the real consume protocol, sealedDealPreimage and parseSealedDeal.

The trace (one test, 6 steps; the artifact is written when PCC_REAL_TRACE_OUT is set)

1. validate: every node current, and the live terms equal the DB rows;
2. the payer issues a reservation;
3. accept: 200, layer B, sealed; the read-back is consumed with the same digest;
4. the stored sealed deal hashes to the digest and parses;
5. a real provider price change in the DB: the same old claim is now stale, with a price diff and the new live price;
6. MC 9: the print operator carves a child reservation from the sealed unit, with its own wallet.

The artifact is /mnt/sparkbulk/pcc-reconciliation/returns/pcc-composition-real-state-trace.json.

What is still a labelled stand-in, and why

• requestTerms and payerFor: gateway-owned, not built (#3503).
• The program resolver and gate: evidence feat(spec): committed programs per CSD tier and the pre-funding assurance gate (must-close 6) #349, not merged.
• evidenceFor: derived from the real CSD JSON tiers, one requirement per primitive, keyed on bind ?? id. Evidence owns the canonical mapping. Keying on id alone collides: this CSD declares machine.execution_log twice, and the compiler refuses duplicate (evidenceTypeId, tier).
• The fee policy.
• The deal encoder: escrow feat(contracts): freeze the V-next settlement ABI + canonical compiler @pcc/contracts/vnext (R14) #367, not merged.

⚠ Gap found: the reference vertical's CSD is not registered in production

loadBuiltinCsds() (packages/spec/src/csd/registry.ts) never registers document-print-and-mail.csd.json, a workflow-kind CSD. So in production, getCsdRegistry().findUrlByType("document-print-and-mail") is undefined: the print leg of the reference vertical cannot resolve its CSD, and validate or accept would answer incompatible.

• The test PROVES the gap first.
• It then registers the unmodified JSON through the registry's own register(), the same call POST /api/csd makes.
• The fix belongs in the builtins list. It has been raised with the steward to route to its owner.

Tests

• This file: 1 test, 6 steps.
• With reservations-route and agent-plans-route in the same run: 28/28. No singleton leaks between files.
• tsc: no errors in this file.

Written by a sonnet subagent (implementer-delta) to the lane's spec, and reviewed and re-run by the lane.

pcc-composition 8a0f4de0, goal pcc-reconciliation.

🤖 Generated with Claude Code