web/programs-react: reconstruct inline virtual activations in the tracer

[gnidan]

Fixes the tracer's rendering of inlined internal calls (the level-2 `inline` transform from #230). Part of #23 (UI side). Targets `transform-context`.

Problem

An inlined body's instructions each carry a call context bearing both `invoke` and `return`. `extractCallInfoFromContext` checks `invoke` before `return`, so every inlined instruction read as an invoke and `buildCallStack` pushed a virtual frame per inline site but never popped — phantom `dbl > dbl` frames accumulated and leaked to the end of the trace, indistinguishable from real calls.

Change (programs-react)

• `CallInfo`/`CallFrame`: add `isInline` (virtual-activation tag).
• `extractCallInfoFromInstruction`: set `isInline` from the `inline` transform marker (alongside the existing `isTailCall`).
• `buildCallStack`: tag virtual frames, and add a defensive per-instruction membership guard — once execution reaches an instruction lacking the `inline` marker, any trailing virtual frame is torn down, so a virtual activation can never leak into caller code.

Change (web + programs-react UI)

• Render an `⧉ inline` chip (dashed, italic; shares the transform palette) on the call-stack frame in both the docs `TraceDrawer` and the standalone `CallStackDisplay`, mirroring the tailcall chip — visually distinct from a real call.

Verification

• TDD: 6 new unit tests in `mockTrace.test.ts` (isInline extraction; virtual frame lifetime across a bracketed body; two gap-separated sites with no accumulation; defensive-guard force-pop; real frames stay `isInline`-falsy). Full programs-react suite green (47).
• End-to-end against real O2 bytecode (`dbl` inlined at 2 sites): each site now shows one virtual `dbl` activation across its body and the stack returns to top level between/after sites — no phantom frames.

Notes

• Robust to the current (pre-de-smear) emission: the consecutive-invoke dedup collapses the smeared per-instruction invokes into one frame and the guard pops it at the caller gap. The compiler-side de-smear (bracket invoke on entry / return on exit) is still tracked under Fix typos in format spec #23 for spec conformance + graceful degradation, but the UI no longer mis-renders regardless.
• Deferred pending architect ruling: display semantics for a genuine single-instruction inlined body (frame visible at that step vs. push-then-pop). Not exercised by current examples.

[github-actions]

PR Preview Action v1.8.1
Preview removed because the pull request was closed.
2026-07-03 00:53 UTC