bench: complex multi-system workflow benchmarks (email + spreadsheet + API + multi-app)

[abrichr]

Why

The existing benchmark suite is strongest on bounded, single-surface workflows. This PR adds two deterministic, synthetic back-office programs that exercise loops, branches, exception terminals, multiple applications, document/spreadsheet intake, governed API-tier actuation, and persisted-state verification.

It also closes four review gaps in the initial benchmark implementation:

1. the governed arm now passes the real Standard-profile run gate and derives its runtime authorization from that gate report;
2. the adjudicators enforce allowed record-level transitions across every non-echo table, catching unrelated inserts, deletes, and in-place edits;
3. expected state comes from immutable fixture seeds/source bytes rather than the prepared worklist produced by the system under test; and
4. CI validates the committed n=3 evidence shape and exact headline, including the correct retry denominator.

What changed

AP invoice processing

• Deterministic vendor email and PDF intake into a five-row worklist.
• Two fixture applications: ERP plus mail gateway.
• A real ProgramGraph with a loop, match/exception branch, and discount branch.
• 32 executed consequential actions on the healthy path.
• Designed faults: missing PO, ambiguous duplicate invoice, collateral adjacent-row overwrite, and uncertain payment confirmation with idempotent retry suppression.
• Separate persisted-state reads: read-only SQLite, payments REST read, and direct OUTBOX maildir read.

O2C reconciliation

• Billing CSV export in, ledger comparison, results CSV write-back.
• Two fixture applications: billing plus ledger.
• A real ProgramGraph with a loop, three-way disposition branch, and explicit missing-record halt terminal.
• 26 executed consequential actions on the healthy path.
• Designed faults: missing ledger record, ambiguous duplicate, stale snapshot, and phantom workbook write.
• Separate persisted-state reads: read-only SQLite plus direct CSV re-read.

Governed admission and adjudication

• benchmark/multiapp-standard.yaml provides the deterministic API-tier certification policy.
• evaluate_run_gate(..., profile=standard) performs certification, identity/effect coverage, verifier-tier, durability, settling, and manifest checks.
• build_runtime_authorization(...) binds the successful admission to the exact sealed bundle and runtime inputs.
• API identity contracts are revalidated by the replayer before request dispatch.
• Direct-file adjudicators never consume HTTP acknowledgements, UI banners, verifier verdicts, or the prepared worklist. They derive expectations from immutable source fixtures and compare exact allowed business-record transitions. They are separate read paths, not separate services or failure domains.

Evidence

Each benchmark runs n=3 for every arm/scenario cell.

benchmark	base runs	extra duplicate-attempt checks	governed healthy VERIFIED	governed silent incorrect success	healthy over-halts	naive silent incorrect success
ap_invoice	30 (15 per arm)	6	3/3	0/15	0/3	3/3 collateral cases
o2c_recon	30 (15 per arm)	0	3/3	0/15	0/3	3/3 phantom-write cases

All runs are localhost-only, use synthetic data, and make zero model calls. This is deterministic fault-coverage evidence, not a customer reliability or production-incidence claim. UI perception is outside these two benchmarks because every action uses the replayer's API tier.

Validation

• pytest -q tests/test_ap_invoice_benchmark.py tests/test_o2c_recon_benchmark.py tests/test_release_contract.py: 57 passed.
• pytest -q tests/test_run_gate.py tests/test_execution_profiles.py tests/test_policy.py tests/test_policy_program.py: 120 passed.
• ruff check and ruff format --check on all changed benchmark/test code: passed.
• scripts/check_release_consistency.py: passed; regenerated public-artifacts.json includes both result files and the benchmark policy.
• scripts/check_consistency.py: passed.
• scripts/validate_claims.py --check: passed.
• Committed n=3 result matrices were regenerated from clean commit c1231aa and committed as 9d132e1.

Scope

• Synthetic fixture code and results only; no customer, PHI, real-EMR, deployment-derived tuning, or private failure corpus data.
• Benchmark files remain repository evidence and are not included in the PyPI wheel.
• PR feat(effects): verifier adapter platform (Section 4) #264 remains open, so this branch has intentionally not been rebased yet.

DO NOT MERGE without review.