[m2kulkarni]

Closes #34. Depends on #33. Author Codex root puffer-full-20261022; independent reviewer Codex review_26, final published-commit recheck pending.

Adds the production Game API for complete simultaneous per-power string order sets, structured observations with all82 locations/actor/unit slots/retreat options, local order candidates, seeded scripted play and hash-verified replay. The retained Discrete1088 interface is explicitly restricted, not the full research action/observation interface.

Fixes actual initial/Fall/reset ownership, nonhome flags, impassable Switzerland, owned empty home/coast builds, mixed Winter sets, support destination reachability, convoy endpoints, typed disbands/retreats, occupied retreat choices, parent-province retreat conflicts, per-unit result mapping independent of submitted order, canonical army coast build/retreat locations, duplicate coast-alias order input, legacy step argument copying and close guards. State/map query reference ownership is corrected. Rules and limits are documented in docs/development/full-order-interface.md; paper Winter welfare/local reference priority retained. Original DATC expectations unchanged.

The old46 skips were audited:26 empty placeholders and several obsolete interface expectations supplied no coverage. They now have executable C/public-game tests; three valid support/circular tests were enabled unchanged. Complete source-pinned standard.map fixture checks every direct army-base/fleet edge (zero differences). New public tests include full support+retreat and convoy games, zero-unit builds, malformed input atomicity, coast collisions and reversed retreat result order, six complete seeded games with per-phase invariants, and replay.

Before fixes new core tests:4 failed/2passed. Each later independently found defect got a failing regression before its fix. Independent reviewer repeatedly reproduced full suites and found the additional bugs above, which are fixed; final exact-head review pending. Final local combined suite is recorded in follow-up after completion. GCC native builds on Linux x86_64/Python3.12.14/Torch2.14.1+cpu. No sanitizer/GPU/performance/strong learned baseline claim. Packaging and installed CLI are separate #30 work. No merge without explicit human approval.

[m2kulkarni]

Final independent review by Codex review_26, recorded by coordinator root. Exact published commit9385d18cb7cad5b25d511345771cae45afb73915 independently fetched/built/tested in a separate checkout: no remaining blocking findings from audited cases. Both GCC native builds passed; combined341 tests and16subtests passed with zero failures/errors/skips. Diff check passed. All identified game/interface defects now have reviewed regressions. Original DATC expectations unchanged; obsolete skipped scaffolds/test corrections are accurately documented. Complete Game API and restricted numeric MLP integration are distinguished honestly. Packaging/GPU/recurrence/fullresume/sanitizer/longrun/performance are outside this change. COMMENT only, no independent GitHub approval or merge permission.