Draft: TC1-Real M3 temporal applicability proof - #19

Stack relationship

This PR is stacked on feat/task-context-tc1-real-v0.1 / PR #18. Do not merge it independently before the TC1-Real M1 direction is accepted.

Goal

Prove the frozen M3 temporal-mismatch case with two real HRRR provider artifacts, while keeping UTC interval reasoning outside GeoTask Core v0.1.

Real A/B evidence

Both HRRR artifacts use the same source, model date, 06Z run, Phoenix task bbox, UGRD/VGRD/VIS, and 10_m_above_ground/surface levels.

control f04 -> valid 2026年08月18日T10:00:00Z -> 594 bytes
mismatch f02 -> valid 2026年08月18日T08:00:00Z -> 596 bytes

The frozen benchmark task window is [2026年08月18日T10:00:00Z, 2026年08月18日T11:00:00Z).

Result

10Z control
 temporal applicability = applicable
 context status = sufficient
 critical gaps = none
08Z real mismatch
 temporal applicability = not applicable
 reason = validity_before_task_window
 context status = insufficient
 critical gaps = weather_wind, weather_visibility

The mismatch weather candidate fails at exactly the existing Core boundary temporal_scope_mismatch. Recorded UASFM/DDOF context is held fixed and retains the same coverage.

The result is the same under network_byte and carried_byte burden projections because cost accounting does not change applicability.

See docs/reference/task-context-tc1-real-m3-result-v0.1.md.

Method boundary

benchmarks/tc1_real/temporal_applicability.py performs the real timestamp/window check. Only an applicable artifact is normalized to the existing opaque recorded-experiment-window scope reference. An out-of-window artifact retains a mismatching scope reference and Core fails closed.

GeoTask Core is not changed to parse datetime values or implement interval algebra in this PR.

CI evidence

Implementation head before the result report:

726974afb38ba9afd06fead2c87023871cfd0005
geotask-core #188 -> completed / success

Report-complete exact head:

72093ad28aa2bc0a1eadfb9c9cfdba64fddf129d
geotask-core #189 -> completed / success

Both combined runs covered:

• Python 3.10 / 3.11 / 3.12 / 3.13 full pytest;
• public boundary/export;
• artifact roundtrip;
• build + twine checks;
• public export scan;
• RC build / Reference Agent replay evidence;
• merged RC evidence readiness.

The PR was temporarily retargeted to main only to trigger the repository's existing PR workflow and is now restored to the stacked base without changing the verified exact head.

Promotion question

M1 exposed spatial containment before opaque-scope normalization; M3 exposes temporal validity before the same boundary. Together they make Spatiotemporal Applicability a credible next abstraction candidate.

This PR does not promote that abstraction into Core. Promotion should require an explicit gate and preferably a second independent domain/use case so a low-altitude benchmark does not define the universal contract by itself.

Non-goals

• no flight authorization or feasibility claim;
• no weather forecast accuracy claim;
• no automatic requirement discovery;
• no automatic provider validity inference;
• no Core datetime/interval semantics;
• no cross-domain generalization claim yet.