Add truthful partial task-only Forecast queries (#85)

[deverman]

Outcome

Refs #85. Add filter.forecast="past-and-today" to the existing list_tasks and get_task_counts surface, plus the equivalent CLI option. This is an explicit partial task-only contract, not an exact native OmniFocus Forecast total. No new model-facing tool or mutation is introduced.

The deduplicated union uses remaining actions with local due/planned dates before next local midnight, defer dates during today, effective flags, or the documented configured Tag.forecastTag (not a hard-coded name). Foundation supplies calendar-day boundaries, including DST. Unavailable scheduled actions are included by default; explicit filters still intersect. Project headers, completed/dropped actions and hidden children are excluded. Both list and count return limitations, and the CLI preserves them too.

See docs/forecast-contract.md for exclusions, visibility-preference differences, inherited-date limitations, and paging across day rollover. The roadmap now identifies #85 as the active slice following completed #92 feature UAT.

Validation impact and acceptance

query, with MCP wire coverage. BridgeClient changes attach the day window and propagate count warnings only; IPC, dispatch, retries and timeout policy are unchanged. This explains narrowing the conservative filename-based transport classification.

Acceptance journey: ask a real MCP client how many items are in Forecast (past and today), receive a dedicated deduplicated count matching the paginated list, and have the assistant explicitly qualify it as partial rather than equating it with native Forecast.

Completed:

• Full Swift Testing suite and release build; new deterministic tests exercise the actual production Bridge JavaScript in JavaScriptCore.
• UTC/Makassar day differences, 23/25-hour DST days, configured tag with a different name, absent configured tag, union/deduplication, unavailable actions, dropped/completed parents, project/tag/ID/inbox intersections, paging, cursor scope, and invalid/missing day windows.
• Direct MCP schema/success/error probes: true, exact, and today are rejected; list/count agree for baseline, inbox, availability, flags and empty searches, including all pages and warning propagation.
• Required script installation, one OmniFocus restart, healthy matching development plugin, and the query-tier live semantic canary. Existing baseline contracts remained green.
• New opt-in Swift live regression test passed for Forecast list/count/warnings and inbox parity.
• Native UI comparison: 8 Past / 0 Today badges versus the explicitly partial union of 58 actions. A native overdue/flagged sample matched; this is not an assertion of exact view parity.
• OpenCode UAT used the candidate release binary: get_task_counts({filter:{forecast:"past-and-today"}}) returned 58 and its answer explicitly retained all limitations. The initial attempt accidentally launched the older command because project config overrode OPENCODE_CONFIG; that run is excluded from acceptance evidence. Runtime OPENCODE_CONFIG_CONTENT corrected the server selection without editing user config.
• Whitespace and local Markdown link checks.
• Targeted Forecast MCP smoke: 610.1 seconds of measured rounds, 28 count/list pairs plus 19 initial semantic calls (75 calls total), zero errors and zero parity failures. Baseline/inbox/availability/flagged scenarios rotated. Median 12.12s and p95 22.41s include MCP round-trip time; no performance improvement or comparative regression conclusion is claimed. Tail latency needs the separate release diagnostics. The final subsequent CLI-only warning projection was independently built/tested/live-probed; MCP/Bridge production code stayed unchanged during this smoke.

The selected Swiftly 6.3.3 compiler failed before package compilation against the installed SDK (-target-arch-variant); verification used Xcode's bundled Swift 6.4 via xcrun/PATH. No toolchain pin was changed.

Release boundary

The realistic 1.5-hour suite is still required on the frozen release candidate before release acceptance. This PR is not a release certificate and does not close #85. #222's release diagnostics remain separate. No Homebrew release, merge, or task write is performed here. Raw local harness artifacts remain under .build/benchmarks; only the selected evidence is published.

Checklist:

• Implementation, automated contracts, release build, installed health/semantic canary, direct wire probes, client UAT, and targeted 10-minute smoke.
• Realistic validation on the frozen release candidate before release acceptance.

Implementation commit: febbaab4734ff9f0fe273a666d364b41d35ef2ab.