feat: --auto-spindump + multi-kind cluster sidecar (HangBuster PR 2/3)

[conorluddy]

Context

Building on PR #91 (which wired --auto-sample to xcrun simctl spawn sample), this PR:

1. Adds a second capture mechanism via xcrun simctl spawn spindump behind --auto-spindump.
2. Fixes a latent gap: auto-samples stashed during capture were never reaching summary.json. The worker writes to auto_samples.jsonl, but SummaryBuilder.build never read it back — so --get-details on a stopped session showed no stack. Only the --resample path (which re-fires sample at inspection time) worked. Bundling the fix here because PR 2 has to touch the same code paths anyway.
3. Generalises the sidecar to multi-kind so both captures coexist under one fingerprint.

This is PR 2 of 3 in the stack-enrichment plan. PR 3 will add atos post-processing on the captured stacks for source-line symbolication.

Changes

Storage (hang_sessions.py):

• read_auto_samples return type: dict[str, dict] → dict[str, list[dict]], preserving write order. Callers disambiguate via the kind field on each payload.
• build_summary now reads auto_samples and passes them to SummaryBuilder.build so they survive into summary.json.

Pipeline (hang_pipeline.py):

• Cluster.auto_samples: list[dict] | None alongside existing auto_sample (kept for backward compat with old summary.json files; rehydrates from either).
• SummaryBuilder.build takes auto_samples_by_fp and attaches matching captures to clusters at build time.
• format_cluster_detail iterates samples, renders each kind under its own header (simctl-sample stack (top 10): / spindump stack (top 10):). Also fixes a latent bug from PR 1: stack[:10] sliced 10 chars when stack became a string; now splits to lines correctly.

Watcher (hang_watcher.py):

• New --auto-spindump CLI flag, parallel to --auto-sample.
• New _attempt_auto_spindump(udid, pid) shelling to xcrun simctl spawn <udid> spindump <pid> 1 -file - with a 10s timeout (spindump is heavier than sample).
• Worker adds a parallel spindumped_fingerprints dedup set so each kind captures at most once per fingerprint.
• --resample populates the new auto_samples list slot.

Tests

8 new units in test_auto_sample.py / test_hang_pipeline.py / test_hang_sessions.py covering:

• spindump subprocess: success / missing-udid / timeout / non-zero exit
• multi-kind storage round-trip under one fingerprint, write-order preserved
• SummaryBuilder.build attaches by matching fingerprint, ignores non-matching
• format_cluster_detail renders both kinds with labels, falls back to legacy auto_sample, reports failure reasons

pytest tests/         153 passed
ruff check            clean
black --check         clean

Verification

• Unit: pytest tests/ — covered.
• Manual smoke: boot a sim, run python scripts/hang_watcher.py --start --auto-sample --auto-spindump --bundle-id <id>, reproduce a hang, then --stop and --get-details --cluster 1 should now show both a simctl-sample stack block and a spindump stack block in the cluster detail. Without --auto-spindump the existing PR 1 behaviour is unchanged (single sample block). (Not run as part of CI — sim-coupled.)