Reduce Cortex placement work

[genixpro]

Cortex placement scans repeatedly traverse the building array and recompute inn clearance for every candidate tile. This change snapshots live building positions, upgrade reservations, and inn occupancy once per placement search, then reuses them for candidate checks.

Scan order, scoring, RNG calls, observation cadence, and corner/wrap behavior are preserved. The snapshot is local to one search and needs no cross-tick invalidation.

Performance and validation

• Eight-team Playground on integration revision c71373bb: 119.77 s → 84.19 s with Cortex alone, about 30% less elapsed time; instructions 2.098T → 1.526T. Replay, initial save, and final autosave are byte-identical.
• Single-run observations, with some concurrent build activity; see docs/cortex-placement-performance.md for scope and caveats.
• Standalone Cortex patch on master: complete SmallForTwo replay matches the original byte for byte.
• Optimized client rebuild and 57,600 geometry comparisons pass; the harness is added to Linux CI.

This PR contains only Cortex placement optimization and its tests/documentation. Save buffering is a separate, independent PR. Both target master; unpublished Maxima integration commits are excluded.

[genixpro]

Rebased onto current master (e3a01b051) via e2268715c to clear the conflict. Three files conflicted, all append/append (an independent new CI step and SConscript test target, and two independent doc sections in test/README.md); kept both additions in each case. The actual production changes (CortexPlacementGeo.{h,cpp}, CortexPlacementCandidates.cpp) and the new geometry harness applied cleanly with no changes needed.

Verified locally before pushing:

• scons -j8 release=1 server=0 cortex-geometry-test builds clean on this head.
• ./build/src/CortexGeometryHarness passes: 57,600 candidate comparisons against the tile-scan helpers.

CI is running on e2268715c. This has had zero reviews so far; requesting a look from @Giszmo.