[Epic] Prove the Flect 1.0 product experience end to end

[robinbraemer]

Outcome

Prove the complete Flect 1.0 experience as one calm loop: open, become agent-ready, describe an interface, see it running, refine it by conversation or direct selection, recover from failure, revisit history, reopen, and export.

This epic is the integrated product-quality proof for #1. It must validate what users actually experience, not a separate review-state machine.

Required slices

• Make first-run Pi sign-in and model readiness zero-friction #19 first-run Pi authentication and automatic model readiness
• Make Flect feel instant with strict live-editing performance budgets #20 strict interaction, build, cancellation, and memory budgets
• Persist workspace, conversation, and canvas state across reloads and failures #21 restart-safe work and browser-storage recovery
• Complete accessibility, reflow, and adaptive appearance gates #22 keyboard, screen-reader, reflow, contrast, motion, and appearance support
• Ship trusted clean-machine macOS distribution #23 trusted clean-machine macOS distribution
• Use Git-backed canonical workspaces for invisible history, undo, and recovery #25 Git-backed history, Undo, Redo, and restore
• Run the canonical Flect quality gate in GitHub CI #31 canonical CI quality gate
• Make the running interface the live editing canvas #32 running interface as the live editing canvas
• Host each Flect app in a persistent incremental frontend workspace #33 persistent incremental frontend workspace
• Give the agent typed tools to inspect and repair the running frontend #34 typed source/runtime inspection and repair tools
• Add element selection and direct manipulation for targeted UI edits #35 element selection and direct manipulation

Acceptance journey

From clean supported browser and macOS profiles, a user can:

1. Open Flect and complete Pi/model readiness without a terminal.
2. Ask for a small interface and see a valid running result.
3. Make ten consecutive UI changes through natural language and selected elements without Keep, Reject, Review, branch, mode, or reset decisions.
4. Continue typing and interacting with the canvas while the agent streams and the workspace rebuilds.
5. Encounter an invalid patch, keep using the last-known-good result, and let the same agent repair it from actionable diagnostics.
6. Undo, redo, inspect a simple history, restore an earlier state, reload, and reopen offline without losing work.
7. Complete the flow with keyboard and assistive technology at supported viewport and appearance settings.
8. Export an ordinary Git-backed web project and run it outside Flect.
9. Install and repeat the supported journey from a trusted clean-machine macOS artifact.

Release gates

• Every referenced issue meets its observable acceptance criteria.
• Production Chromium proves the full journey against production assets and real browser behavior.
• The packaged macOS application proves the same supported journey through public UI and accessibility surfaces.
• Performance budgets from Make Flect feel instant with strict live-editing performance budgets #20 pass without excluding the agent-active, failing-build, long-session, or direct-manipulation paths.
• Capability enforcement, credential isolation, redaction, revocation, storage recovery, and last-known-good behavior remain deterministic.
• A dated verification report links current evidence for every advertised claim and names every known limitation.
• The user-facing product contains no internal App/Shaper/Guardian roles or mandatory candidate/review terminology.

Non-goals

• Treating an old Shape/Use/Keep workflow as a compatibility requirement.
• Claiming the later remote-runtime exploration is shipped.
• Passing through mocks alone when a production browser or packaged-host boundary can be exercised.