# Shared gateway acceptance coverage

`cases.json` preserves all 47 named scenarios from sections 7.1–7.4 of the unified gateway plan. These are requirements, not recorded passes. Native deployment/topology and performance requirements in section 7.5 remain additional gates; this manifest does not claim to enumerate their combinations.

The implemented checker rejects missing/extra/duplicate scenarios, non-passing outcomes, missing assertions/evidence, wrong OS/SHA, non-production dispatch, and leaked owned processes. It checks evidence availability, not whether a trace proves its claimed observation. Human/independent trace review remains mandatory. A loopback model is not live-provider evidence.

Run against a real driver's report:

    python3 evals/shared_gateway/verify_receipt.py receipt.json --expected-sha FULL_TESTED_SHA --expected-os linux

Evidence paths are relative to the receipt directory. Supply a full report with schema=1, source_sha, native_os, fixture_home_isolated=true, production_dispatch=true, model_boundary (loopback or live-provider), remaining_owned_processes=[], and cases. Each case contains name, status, positive integer assertions, and nonempty evidence_paths pointing to nonempty files.

## Retained standalone reader A/B

`search_readonly_ab.py` runs the real registered session-search tool in fresh interpreters with disposable homes. Its three cases cover empty, foreign-profile and existing-profile reads on both trees. It asserts the baseline creates an unwanted current-profile DB or commits a read-side write, while the fixed tree preserves marker reads without those mutations. Run with the project venv Python and explicit `--baseline`, `--fixed`, and private `--output` paths. The script records both tested SHAs and search-source hashes. Its receipt is deliberately narrower than the full `readonly-client` scenario, which also requires list/export and other reader paths.

`session_readers_cli_ab.py` exercises the real CLI entrypoint for insights, sessions list, stats, and pinned JSON. It runs each command with both empty and populated disposable profiles on each tree, verifies successful output, and compares database hashes before and after. The selected historical baseline must exhibit read-side writes; the fixed tree must not. It does not cover exports, native interactive clients, or concurrent-owner reads.

    python3 evals/shared_gateway/session_readers_cli_ab.py --baseline /path/to/baseline --fixed /path/to/fixed --python /path/to/project/.venv/bin/python --output /private/path/readers-cli-ab.json

## Normal gateway bootstrap descriptor probe

`normal_boot.py` launches `python -m gateway.run` with a disposable, credential-free profile on native POSIX. It polls the real control socket until the deadline and exits nonzero unless the owner advertises `ready`, protocol 1, an authority epoch, and an API origin. The private receipt retains descriptor transitions and the child shutdown result; the adjacent log retains startup diagnostics. Only its owned process group receives cleanup signals. An unexpected SIGTERM normally gives the gateway exit code 1; this is separate from the probe's readiness verdict.

    python3 evals/shared_gateway/normal_boot.py --repo /path/to/checkout --python /path/to/project/.venv/bin/python --output /private/path/boot-before.json

Repeat against the fixed checkout with a different output path. A descriptor pass is **not** proof of an authenticated API handshake, session execution, worker ownership, or native Windows startup. The integrated bootstrap is expected to pass it (`tests/gateway/test_normal_runtime_boot.py` asserts the same `ready` descriptor from the real control socket); a failure is a regression, so retain it rather than weakening the assertion.

The unified `run.py` driver is not implemented yet. Existing private native and subprocess probes have different receipt formats; do not rename them or manufacture a full report to obtain a pass. Test fixtures in `test_shared_gateway_harness.py` exercise the checker only and are not acceptance evidence.