Draft: prove task-appropriate provider representation

[stpku]

Stack relationship

This PR is stacked on feat/task-context-provider-truncation-proof-v0.1 / PR #30. It was temporarily based on main for combined exact-head CI and has now been restored to the stacked base.

Goal

Turn the second Phoenix planning failure into a small offline proof:

Even after provider completeness is proven, the complete record count is not automatically a valid Task Context baseline if the chosen source representation duplicates the physical-world entities required by the task.

Recorded real evidence

Joined Growth layer 2 (FinalLUAUs_PopEmp) was acquired completely:

complete features        118,190
unique planning units          471
network bytes        172,127,458
pages                         119

The same frozen broad region using base planning-unit layer 4 (FinalLUAUs) recorded:

complete base units           471
unique planning units         471
network bytes              24,237
requests                        2

Population semantics are related separately through table 13 using newluau.

The joined representation therefore produces about 250.9 feature records per unique planning unit.

Anti-overclaim

The 172 MB vs 24 KB byte contrast is deliberately non-scoreable. The two representations do not carry identical semantics/fields, so this PR does not claim a 7,000x compression gain.

The supported conclusion is narrower:

Context cost can be dominated by choosing the wrong physical-world representation before task-scope reduction is even considered.

Implemented

• compact recorded representation-evidence fixture;
• dependency-free offline analyzer;
• fail-closed tests for completeness and entity-set parity;
• method report under docs/reference/.

Exact-head validation

Exact head 636537d52626cf9f1775985290f45da92011bdc2 passed combined main CI run #243:

workflow  geotask-core
status    completed / success

Core impact

GeoTask Core code delta      0
public schema delta          0
live provider dependency     0

This is benchmark/method evidence only.