fix: record lost a mark born during the block, and the backtrace of what threw

[sotashimozono]

Two defects in record, both surfaced by reviewing #24 and both older than it. The first breaks
the one thing the default layer promises.

A mark born while the block ran was lost — from the always-on layer too

record snapshotted the probe set before calling f and never looked again. A probe that came
into existence during the call was entered by code that ran, counted by nobody, and left with its
flag false for the rest of the process — because while a recording is open the write side
counts into the probe instead of setting the flag, and only record's epilogue sets it, over the
stale snapshot.

before:  record saw [:tracked]              entered(D) = [:tracked]
after:   record saw [:newborn, :tracked]    entered(D) = [:newborn, :tracked]

newborn had run and returned a value in both cases.

A package extension loaded inside the block is the ordinary way this happens, and this package
ships three of them. So entered() and the exit summary — the layer whose whole claim is that it
cannot miss an entry — would have said nothing.

The probe set is now re-derived after the call, saved is keyed by probe rather than by position,
and reconciliation runs over the union. Re-deriving needs invokelatest: the new probes' bindings
are younger than the frame reading them, and 1.12 warns that will become an error.

The exception's backtrace pointed at record, not at the caller

throw(err) after the catch block manufactures a fresh backtrace:

via record, before:  ["error", "record", "top-level scope", …]
via record, after:   ["error", "energy", "deep", "mid", "outer", …]
direct call:         ["error", "energy", "deep", "mid", "outer", …]

A user debugging a failed run saw record.jl where their own call chain should be, with nothing
to say frames had been dropped — in exactly the case the docstring names as the reason to reach
for rethrow = false. Closing now happens inside the catch and the exception is re-raised with
rethrow(), which keeps the backtrace it arrived with.

The comment that claimed this was impossible was right about Base.rethrow(err) outside a catch
and wrong about the conclusion: the call had simply drifted to after the catch.

Tests

Both fail against the unfixed file, and each carries a control — a mark defined and never called
is still absent (so the fix did not start reporting everything it can see), and the direct call is
shown to carry the frames the recorded one must also carry.

1062 assertions, green.

🤖 Generated with Claude Code

[github-actions]

📚 Docs preview: https://codes.sota-shimozono.com/ExperimentalAPI.jl/previews/PR25/

(updates on each push to this PR)

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!