feat(flow): parallel worklists with one session per worker, measured

[abrichr]

An outside review said OpenAdapt's throughput is serial and no faster than careful human entry. This PR makes worklists parallel, makes one entry faster without removing a check, and records where every step's time goes. Speed figures are published only by the scoreboard (benchmark/scoreboard, launch-demo session): this PR provides its arm 4 (fast speed profile) and arm 6 (--workers N).

What changes

Parallel worklists

• replay and run take --workers N for a for-each bundle (one loop over a worklist). Each item is the same governed run a serial loop makes for that row, with every check, in its own run directory with its own one-row authorization.
• Each worker signs in once with its own account from the deployment file's new worklist: section (env var names only). The run refuses N above worklist.max_workers, too few accounts, two accounts for the same user, two workers sharing a session (same cookies), rows not declared items_independent, and programs with steps outside the loop.
• Items are claimed once from a durable SQLite queue in the run directory. An item claimed before a crash becomes uncertain and is never re-run automatically.
• A halt, failure, worker error or RECONCILIATION_REQUIRED on any worker stops new claims on all workers; in-flight items finish their own checks. worklist_pool.json and WORKLIST.md show per-worker and per-item results.
• --readback-window K: a worker may start item n+1 while item n's system-of-record read-back runs (only the item's last action, only read-backs that don't use the screen, never for durable runs). The item report is rewritten with the verdict. A failed read-back stops all workers and names the items written after it.
• worklist.entry_bundle: a recorded navigation replayed before each item, for apps that keep state between items.
• ProcessWorker: the same worker in its own process (same queue and stop flag). RDP workers use the existing FreeRDP SessionPool (unit tests only).

Where the time goes

• Every step in report.json has timings: exclusive milliseconds for settle_before, screenshot, resolve, identity, ocr, actuate, settle_after, postcondition, readback, artifact_write, other. They add up to elapsed_ms.

One entry, faster (every check kept)

Always on (no tradeoff found): scroll readiness asks the DOM first and keeps scrolling while the target exists off-screen, without a full-frame OCR search; DOM text postconditions match letters and digits and accept rendered text when hidden option lists interleave textContent (a miss still falls back to OCR); a structural lookup checks box and frame chain once, after binding.

Named speed profiles (openadapt_flow.runtime.speed, runtime.speed), default standard:

Setting	standard	fast	Tradeoff
Typing into plain text fields	key by key	one insert (input, then change on blur, fire)	Per-key handlers in a plain text field don't run; comboboxes/autocomplete/datalist and the guarded keyboard path always get every key; input read-back and DB read-back still check the value
Page-quiet window	100 ms	50 ms	Less time for the app to start reacting before the next check; a later reaction becomes a halt, not a wait
OCR threads	default	2	Helps many workers, can slow one

Not touched: the guarded click's re-checks, identity checks, input read-back, the system-of-record read-back.

Found along the way

• The identity check refuses parameter values under 4 characters (first name "Bo" halts). Now in LIMITS.md.
• The final confirm button's OCR identity band includes form text behind the dialog, so it can halt when the page scrolled a little differently. Fails closed.
• The referral bundle halts at a scroll step in installed Google Chrome (the default auto browser); bundled Chromium works.
• Transaction outcomes are upper case in reports; the pool now compares case-insensitively.

Still open

• Root CAPABILITIES.md rows wait for docs: one map of what OpenAdapt can do #1260, which adds that file.
• Timed runs and figures: the scoreboard, only while host load is under 12.

Tests

test_worklist_pool.py (claims, a forced halt on one worker stopping the others, refusals, resume, pipelined window, process workers), test_readback_pipeline.py (real Replayer against MockMed), test_cli_worklist_pool.py (CLI wiring, refusals, entry bundle), test_fast_path.py (native fill, speed profiles, off-screen probe, hidden option lists, step timings), test_page_settle.py (punctuation case).

🤖 Generated with Claude Code