Help consumer AI agents Discover/Confirm with complete RunVars

[ds-code-t]

Summary

Dan asked for four related 2.1.10 improvements so consumer AI agents run Discover/Confirm faster and stop being confused by RunVar inheritance.

• Discover/Confirm mvn test now has an explicit complete pkb_runvars: pkb_browser=CHROME_HEADLESS, a conservative high pkb_parallel, pkb_reportingmode=diagnostic, pkb_loglevel=warn, pkb_reportretention=failed. Isolate / live Workbench stays one paused scenario.
• DiagnosticCli discover-hint prints the estimated integer parallel count (not 80 and not a placeholder). pkb_parallel=auto resolves at run start and the integer is stamped into pkb_run_profile.
• Pickleball ships a bundled CHROME_HEADLESS yaml. Consumer pkb_configpath / configs mapping still wins when present.
• No new public input key. Agents must set complete Discover pkb_runvars, then read pkb_run_profile from catalog / run-index / summary. Workbench diagnostic MCP tools surface that field.

Parallel formula

max(2, min(availableProcessors, floor(maxMemoryMB / 512), 24)) from Runtime.availableProcessors() and Runtime.maxMemory() only. Cap is 24 so a 32-core / 64GiB box does not blindly start 32 Chrome workers.

CHROME_HEADLESS fallback

ParsingMap.initializeConfigs loads the consumer config mapping first. If CHROME_HEADLESS is already present (any case), that local yaml wins. Otherwise Pickleball injects META-INF/pickleball/configs/CHROME_HEADLESS.yaml (--headless=new, fixed window size, no MAXIMIZE, QUIT_LOCAL_DRIVER). CHROME / EDGE / GRID / SAUCE consumer configs are unchanged.

Complete snapshot

pkb_runvars remains the only controlled input. After a diagnostic run, run-catalog.json, run-index.json, and summary.json all carry runProfile. Workbench workbench_diagnostic_* tools return that JSON. The worker still resolves internally through PickleballRunner and still rejects pkb_run_profile as input.

Functionality-change coverage

• I reviewed AGENTS.md and the relevant feature-map entry.
• Framework implementation is complete.
• Focused framework tests were added or updated.
• Maven consumer scenarios and supporting resources were added or updated when applicable.
• README or canonical documentation was updated when behavior changed.
• docs/agent/feature-map.md remains accurate.
• docs/agent/repository-index.md is current.
• Backward compatibility was preserved, or the breaking change is documented.

Pointer files maven-consumer-project/AGENTS.md and .github/copilot-instructions.md stay short. No version bump. No GUI Workbench work. Did not change agent-pointer-eval.feature.

Validation

• python3 scripts/verify_agent_contract.py
• python3 scripts/sync_consumer_guidance.py --check
• python3 scripts/refresh_agent_index.py --check
• ./gradlew test (BUILD SUCCESSFUL)
• ./gradlew publishToMavenLocal
• Focused Maven consumer: pkb_browser=CHROME_HEADLESS, pkb_tags=@diagnostic-single or @diagnostic-reporting — 2 scenarios passed, including new parallel/Chrome/discover-hint/catalog runProfile checks
• @agent-pointer-eval is not on @all / @regression (contract check + feature tags). Default @all does not start that harness. Did not "fix" agent-pointer-eval.feature.

Did not run full @all.