fix(server): stop emitting an empty interrupt frame at natural end of…

[decrypto21]

What

A subscription stream that ends naturally no longer emits a zero-length interrupt frame. Natural end of stream emits nothing; only an explicit interrupt carries a terminal frame.

Slot cleanup is unchanged: the subscription task still removes its own generation's slot, and a superseding reservation still owns its own cleanup.

Why (this is D6 from #271)

No deployed host signals natural end of stream on the wire - the novasama host-container sends receive frames only. The core's extra empty interrupt frame is a conformance divergence, and the generated TS client treats the zero-length payload as an undecodable interrupt - so at best it is noise, at worst a decode error surfaced to the product.

This is decision-independent of the error-shape direction (#271 D4 → #259): it is about emitting a frame at all at natural end, not the shape of error frames. It previously rode the flat-wire branch (#285); with that PR closed in favour of #259, this ships on its own - same rationale as the novasama terminal-frames fix (paritytech/triangle-js-sdks#228), which is the mirror image (host answers a frame on every fault; core stays silent on natural end).

Evidence

• truapi-server unit tests updated in place: register_completion_emits_no_terminal_frame (was register_completion_emits_interrupt), supersede/stop/dispose cases assert no stray frames.
• Full cargo test -p truapi-server green (310 tests).
• Differential probe (core vs novasama at triangle-js-sdks 592dc4b): S1 stream-end now "receive only" on both sides (W1-Convergence-Run-2026-07-16.md, row S1).

Relates to #271 (D6).