[codeministry]

Spec 028 takes on the follow-ups deferred from spec 027's code review.

What changes

• The warm-up costs the chat's day nothing until the chat is used. A data change before the first open since start prepares nothing (ISC-536). After that, a change starts the warm-up only once leadgen.chat.suggestion-quiet-window (default 30 s) has passed without a further change, so triaging applications one by one costs one call per pause (ISC-537).
• The quiet window no longer waits behind a scheduled ingest pass. It is timed by CompletableFuture.delayedExecutor, not by Boot's taskScheduler, whose single thread a pass holds for as long as it runs.
• A failed phrasing is retried after leadgen.chat.suggestion-retry-after (default 5 min) instead of waiting for the data to move (ISC-535). At its bound the skip list drops expired entries first, then the oldest, and is never cleared whole.
• A miss is phrased in its own language with the job its request built (ISC-538).
• Follow-ups are no longer stored (ISC-539). They are kept in a memory map of their own, so they cannot push the suggestions out of the table or the cache. The follow-up path now logs a named model that cannot be built.
• event.DataChanged replaces chat.suggest.SuggestionDataChanged. analytics, application and archive import nothing of the chat, and ChatIsolationTest checks imports and every chat_* table (ISC-541).
• The warm-up runs on Boot's applicationTaskExecutor and drops a trigger after shutdown quietly (ISC-540).

No migration, and no value to set.

Verification

• ./gradlew build: 1304 backend cases, 96 unit and 22 browser test files, all green; the 13 skipped are the corpus tests without .eml samples.
• New tests were red before their fix: the busy-scheduler case in SuggestionPrewarmTest, and PhrasingWithoutAContextTest (skip-list bound, follow-up cache, unbuildable-model log).
• Prewarm burst tests use a 1 s window and skip (assumeThat) when the runner leaves a gap of most of a window between two moves.
• Code review in three rounds, security review clean. Ten findings are accepted and listed in the master's Remaining Work.