[abrichr]

What

• Rewrites packages/flow/docs/BACKEND_PARITY.md as a matrix of twelve capabilities by six surfaces: web, Windows, macOS, Linux, RDP and Citrix.
• Every cell is QUALIFIED, TESTED, CODE ONLY or MISSING, with its evidence path. Fakes and stand-ins count as TESTED, never QUALIFIED.
• The page keeps the URL-flags table that tests/test_backend_parity.py reads.
• It links the twelve gap issues, labelled backend-parity and ordered by value to a customer running Windows desktop apps over RDP or Citrix: Live self-repair and teaching a fix over MCP refuse every non-web run #1694 to macOS has no record-compile-run result on a real app #1705.

Findings

• Real third-party apps have run only in a browser (OpenEMR, Odoo, Frappe Lending, openIMIS, Medplum) and over RDP.
• The RDP run is benchmark/rdp_avc_real: OpenEMR in Edge inside a real Windows Server 2022 session. It used a scripted Recorder and Replayer.run, not openadapt-flow run.
• No third-party Windows desktop app has run on any surface.
• No real Citrix session has been run.
• Self-repair, rollback and the agent guard refuse or don't exist off the web.

Claims corrected

Where	Was	Now
CAPABILITIES.md, Surfaces	Citrix in_session: "Same evidence as the Windows path"	No run inside a Citrix session
CAPABILITIES.md, Surfaces	Web: "Real app, every CI build"	Real apps run locally; CI runs in-tree apps
CAPABILITIES.md, Surfaces	RDP and macOS evidence	Names the apps, the scripted harness, and that the macOS trials drove the backend directly
CAPABILITIES.md, Surfaces	Page structure first: "Real app"	Real app on the web only
CAPABILITIES.md and README	Repair and rollback on every surface	Repair and rollback rows say web runs
README	"record, compile, run and repair work on every surface"	Corrected; the Linux and RDP rows in the surfaces table are corrected too
SURFACES.md, LIMITS.md, desktop/LINUX_NATIVE.md	A "required" GTK3 AT-SPI CI job	That job lives in packages/flow/.github/workflows/ci.yml, which the monorepo never runs
desktop/CITRIX_PIXEL.md	win32 client "mock-contract-tested only"	It ran in rdp_avc_real
RECORDING_ON_WINDOWS_RDP_CITRIX.md	"What has been tested"	Updated
rdp_avc_real/README.md	"What this proves"	Says the harness bypassed the run gate

llms-full.txt is rebuilt.

Checks

• tests/test_backend_parity.py, test_backend_parity_refusals.py, test_lookup_capabilities.py, test_validate_claims.py, test_ci_workflow_contract.py, test_backend_capability_matrix.py: 100 passed.
• Root README and copy tests: 22 passed.

Adversarial review

A separate reviewer, given the code and evidence but not the conclusions, checked every cell. Its fixes are in 73f1374:

• Rollback is split from undo. Rollback is refused off the web; undo propose and undo run are CODE ONLY there.
• The surface-neutral guarded save is now in the agent-guard row.
• One rule now covers scripted recordings.
• Windows record and compile are marked real-VM, from windows_uia. Windows replay now shows the full tally: 4 of 10 runs accepted overall, 3 of 3 on the final candidate.
• Shadow mode has the same verdict in every column.
• Web approvals drop to TESTED: the person was scripted, and Cloud was a local stand-in.
• The self-repair row separates live repair (MISSING) from offline teach (CODE ONLY).
• Four refusals were added.
• The Aardwolf RDP batch is now described as a Windows 11 VM.
• The never-run job is named precisely: flow's linux-atspi-x11, not capture's AT-SPI read.