[dayfine]

Base: frlg/stage7-celadon (#224). #260 has merged there; its authored loader and shared codec are inherited, outside this diff. This follow-up contains only the broader cross-repository replay diagnostic and an unchanged route-formatting extraction; the loader/data/equivalence checks are in merged #260. No console or production executor behavior changes.

Load the evaluated Parcel checkpoint and run PokemonSimulation #73's seven scenarios through the existing checkpoint engine and native route runner. Semantic action assertions check complete straight walks, facing changes, map transitions and scenes. Ten native negative controls reject invalid pins/bindings/policies/routes and catch missing, extra, reordered or incorrect actions.

The supplied route is independently generated from the Python planner's actual output, never from test expectations. Native movement rules validate it; the native story-cost planner's own alternative is exported separately. A no-interruption OverworldSimWalker control additionally checks holds, fades and connection stalls. Raw engine initialization and walker diagnostics remain visible in the trace.

This is diagnostic-profile conformance, not live-executor parity: scenes/encounters are scripted, the test policy is stop/no-save, production runtime/save/retry options remain unbound, and native STEP/scene fallback behavior is unchanged. The fixtures pin an explicitly labeled preview of #73; this does not adopt unreviewed production plans or alter the approved map resource pin.

How to review: start with devtools/story-replay/README.md and the linked readable shared cases, then PokemonFRLG_StoryReplay_IO.*, PokemonFRLG_StoryReplay_Executor.* and adjacent tests. The route-formatting move is a separate preparation commit; its two functions are unchanged. Inspect moves with:

git diff --color-moved=dimmed-zebra --color-moved-ws=allow-indentation-change origin/frlg/stage7-celadon...frlg/checkpoint-replay

Local validation: eight C++ translation units pass syntax checks, code health against the intended parent has zero failures, fixture bytes and independent Python candidate regeneration match, and CRLF-aware whitespace checks pass. Initial CI found follow-up issues: Windows rejects two missing designated field initializers; Linux builds and completes the successful Parcel scenario/timed control, but failure scenarios expose an entry-failure label mismatch and empty receipt objects serialized as null. These remain unresolved in this draft; this is not ready to merge. Ubuntu uploads the compiled trace. No console/profile/save was used.

Parent sync at 518b44972: retains the merged authored loader, parser/writer split, shared codec, and std::string expectation. Replay registration follows the moved Data/Json_Tests.cpp, and replay IO includes Data/JsonParser.h. Code health against the new base passes. The existing diagnostic failures above remain unresolved; this sync does not make the draft merge-ready. All 10 affected translation units pass local syntax checks. Fresh Linux native CI is running.

Story action table (core stage 3, from #213/#222): ReplayExecutor now runs steps through the same FRLG rows, binding ROUTE, STEP and MASH to its scripted actions and everything else to unsupported, as before; its guard rethrows, so a replay still never reloads. Gate: the FRLG full-run sim (00.01 to 00.27) step trace and all 12 replay reports are byte-identical to the pre-change baseline, StoryReplay 17/17, and the failing FRLG/Pokemon test names (missing CommandLineTests images only) are unchanged.

Merged #222's core stage 8 through #224: the FRLG sim now runs through the same action table as the console and ReplayExecutor; replay reports and the full-run trace are unchanged.

Interrupts between steps in the replay: ReplayExecutor binds the FRLG table to an optional scenario field between_steps (dialog, wild encounter, trainer sight, warp), each cleared one recorded as an "interrupt" action. cases.json stays PokemonSimulation's copy, so the new test is a mutation of parcel-all-events (StoryReplay 18/18); the 12 existing reports are byte-identical.