[joelteply]

Serves the minds' speed (plan path step 1, goal 10): the learned tool head (about 5k tokens instead of about 10k of schemas, most never called) reaches the engine only if accepts_tail_tools is true, and every resident's binding goes through PlacementSwitch, which didn't forward it. So on the M5 after #5491 deployed, every epoch still read delib.tool_head.epoch sent=false (Aiko: head_tokens=5047 waste_tokens=4514, engine b10995-1c6681593 serving).

Change: PlacementSwitch::accepts_tail_tools asks current(), like every other capability it forwards (capabilities, served_model_ids, live_served_window). A remote seat (AircRemoteInferenceAdapter) keeps the trait default of false.

Test: the_switch_asks_the_seat_she_routes_to_whether_it_takes_her_tool_head, false while routed to the remote seat and true at a home lane that takes the tail. cargo test -p continuum-mind --lib placement_switch: 24 passed (M5, serving down, 00:32Z).

Follow-up for review: a wrapper adapter can still silently miss a new defaulted trait method. Worth considering a default that delegates through generation_target() so wrappers are right by construction.

Measured (after deploy)

Owner Fable: delib.tool_head.epoch sent=true on M5 residents; one real tail-verb call; input tokens/act and s/act vs the 23:31-00:30Z baseline over 15 minutes.

🤖 Generated with Claude Code

https://claude.ai/code/session_01LoTjvf5j3Ez13g6k8mRkFo

[joelteply]

Cormac at b263f0a: source PASS, on CI green. CI was still running when I read it: 5 pass, 5 pending, 4 skipped.

Verified by reading:

• The call is made on the switch. tool_surface.rs:80 asks binding.adapter.accepts_tail_tools(...), and binding.adapter is the PlacementSwitch. Before this PR that hit the trait default (inference_device.rs:421, false), which matches delib.tool_head.epoch sent=false.
• The forward matches its siblings. self.current().accepts_tail_tools(model) follows the same pattern as live_served_window directly above it. The test covers both sides of a placement move: false while she is routed to the remote seat, true after FallHome.
• The remaining race is already handled. The surface decides the tail on current(), while request.rs resolves the dispatch target later. A move between the two would send a tail to a seat that won't take it. That case is already typed: the engine raises InferenceError with "does not take tail_tools now: rebuild the request without its tail" (inference_error.rs:85, 204). So it fails loudly and rebuilds; nothing is silently mis-sent. No change needed.