feat(web-api-discovery): name each post-processing stage and heartbeat the AI pass (#492)

[MarkMichaelis]

What

After a recording stops, HAR post-processing ran for minutes and printed nothing -- ~95% of that time in the spawned claude -p catalogue pass. The operator could not tell a working run from a hang.

• Each post-processing stage is now named as it starts: scrubbing (with the size), verifying the scrub, building the digest, cataloguing.
• The AI catalogue child now runs under a small wrapper that prints an elapsed-time line every 30 s naming the child's pid, and a closing line when it ends. The child's own output still goes straight to the terminal; its exit status passes through (a signal-killed child reports 128 + signal number, so it stays distinguishable from the CLI exiting 1).

Heartbeat rather than streaming, as the issue recommends; streaming stays a separate question. Post-processing stays synchronous -- the wrapper is the process that can keep a timer while the recorder blocks on the child.

Evidence

Same fixture, before -> after: nothing printed -> three stage lines. Heartbeat at its real interval against a 65 s silent child:

[  0.2s] child: working (silent under -p, like the real catalogue pass)
[ 30.1s] capture-har: cataloguing ... 30s elapsed (pid 55064)
[ 60.1s] capture-har: cataloguing ... 1m00s elapsed (pid 55064)
[ 65.2s] capture-har: cataloguing finished after 1m05s
exit: 0

Tests

• New capture: post-processing prints nothing for minutes, so a working run is indistinguishable from a hang #492 section: stages announced before they run and in order; digest announced before it is written; the real pipeline launches the AI child through the wrapper (stdio inherited, cwd = output path, announced first); elapsed-time format; heartbeat names the child pid; heartbeat stops with the child; exit status passes through; signal deaths map to 128 + signal; a child that cannot start is reported, not waited on; the script door passes status through.
• Red first; ablations: removing the timer cleanup and bypassing the wrapper each fail their test.
• Local CI (GitHub Actions is billing-blocked): capture node suites 82 + 16 + 29 + 18 + 14 + store + 13, all green; capture Pester wrappers + node-test-coverage 153 passed, 0 failed.

Independent review

Reviewer model: Claude Sonnet 5 (author: Opus 5), two rounds.

• Round 1: no Critical; 2 Important accepted and fixed (real launch path untested -> injectable launcher + wiring test; signal deaths collapsed to exit 1 -> 128 + signum), 1 Minor accepted (spurious "could not start" after a real exit), 1 Minor rejected (a few hundred characters less command-line headroom; same single process hop).
• Round 2: fixes confirmed; no new Critical or Important findings.

Closes #492
Refs #491

🤖 Generated with Claude Code

https://claude.ai/code/session_01CUt6s3N3iqMMEtvd9AoAnU

[MarkMichaelis]

Evidence -- #492: post-processing says what it is doing

Change type: CLI behaviour (bug fix: silence indistinguishable from a hang).

1. Stage labels -- same fixture, before and after

Synthetic 400-entry capture (166.5 KB), run through the real postProcess
(real scrub, real leak gate). Timestamps are seconds since the call.

Before (origin/main):

(postProcess returned: errors=0, verified=true, catalogue delegated to agent)

Nothing at all is printed while the stages run.

After (this branch):

[0.0s] capture-har: scrubbing 166.5 KB ...
[1.2s] capture-har: verifying the scrub ...
[1.4s] capture-har: building the digest ...
(postProcess returned: errors=0, verified=true, catalogue delegated to agent)

Each stage is named as it starts. The catalogue label is not shown here
because this run delegated cataloguing to the calling agent -- no AI child
was spawned, which is correct.

2. The heartbeat, at its real 30-second interval

The wrapper the AI catalogue child now runs under, driving a stand-in child
that prints one line and then stays silent for 65 s (as claude -p does).
Left column is wall-clock seconds.

[  0.2s] child: working (silent under -p, like the real catalogue pass)
[ 30.1s] capture-har: cataloguing ... 30s elapsed (pid 55064)
[ 60.1s] capture-har: cataloguing ... 1m00s elapsed (pid 55064)
[ 65.2s] capture-har: cataloguing finished after 1m05s
exit: 0

The child's own output still reaches the terminal directly, the heartbeat
names the child's pid, and the wait ends with a closing line.

Review

• A. Does the artifact confirm the intent? Yes -- every stage is
  announced before it runs, and the silent AI stage now shows an
  elapsed-time heartbeat with the child's pid, matching the issue's
  proposed output.
• B. Any new problem? None seen: no stdout output was added (labels are on
  stderr), results and exit status are unchanged, the heartbeat stops
  when the child ends.