A same-model re-select is read from the transcript

ADR 0027 wrote down that modelsAreEqual in pi-ai compares only id and provider, so pi emits no model_select for a re-set to the model the session already holds. That made the re-select the one selection event the extension could not see, and /llama-window the only way to force a compare after it. The operator's case showed the cost in daily use: a llama.cpp model whose live n_ctx no longer matched the window the session held (the report: 160k on the server, 130k in the session), and /model used to pick that same model again did nothing at all. We decided: the extension reads the re-select from the transcript. pi appends a model_change entry on every setModel, equal model or not, so that entry is the record that a selection happened. The wiring keeps the id of the newest model_change entry it has accounted for - a select pi reported through model_select, or a re-apply this extension made itself - and treats a newer entry at the next turn_start as the operator's re-select of the current model. It re-arms both Attempts through a new core entry point, onReSelect, before the pre-request compare runs, so the first request of that turn carries the healed window. The wiring accounts for the entry its own re-apply appends the moment pi.setModel resolves, so a heal can never re-arm the budget it just spent.

Considered options

• Transcript detection (chosen). It needs no change outside this repo, works in every pi mode, and rests on the one fact pi guarantees: setModel appends the entry before it decides whether to emit the event. Rejected cost: the earliest boundary that can read the entry is the next turn_start, so the heal lands one boundary later than the selection and the report line appears with the next turn, not at the moment of the re-select.
• Change pi so _emitModelSelect fires for an equal model. It would re-arm through the existing model_select handler and report at selection time. Rejected because it changes what model_select means for every extension, for the benefit of one, and because ADR 0027 already chose that this extension carries the burden of a stale cached window rather than pushing it upstream.
• Treat any new model_change entry as a re-select, with no self-guard. Rejected because the extension's own re-apply writes exactly that entry. It would re-arm the Attempt the heal just spent and run a compare every turn forever, which is the loop the one-shot budget exists to prevent. The e2e asserts the guard.
• Keep /llama-window as the only recovery. Rejected because it asks the operator to know the internal reason their obvious recovery did nothing. The command stays, for the case it actually covers: a window moved with no re-select at all.

Consequences

• A re-select is a selection event like any other, so the budget rule from ADR 0027 is unchanged: two Attempts per selection, one per Heal moment, and a re-select re-arms both. It cannot loop, because a re-select spends what it re-armed.
• The heal for a re-select lands at the next turn_start, not at selection time. The operator sees the line at the start of the next turn, and that turn's request is already built on the healed window.
• The wiring now reads the session transcript (sessionManager.getEntries) at every turn_start, the same pattern the model router uses. It scans backward for the newest model_change entry and stops at the first hit.
• ADR 0027's sentence that a re-select "re-arms nothing" holds only for the model_select event. The selection itself reaches the core through the transcript.