Improve js-memory leak signal scoping and confidence

[V3RON]

Summary

• agent-cdp users can now run memory usage leak-signal --since SAMPLE_ID to evaluate one bounded interaction instead of accidentally scoring against older daemon samples.
• The leak check now reports confidence, scope, and quality notes so mixed history, missing GC checkpoints, and too-few samples read as weak evidence instead of looking conclusive.
• The heuristic now treats post-GC retained growth as the primary signal and keeps directing users to heap snapshots for confirmation.

Context

The js-memory leak signal previously scored against the full stored sample history and could produce a strong-looking result from mixed workflows. This change keeps the existing sampling workflow, adds an opt-in bounded analysis window, and reworks the scoring/output so retained post-GC growth carries the most weight while trend-only evidence is explicitly downgraded.

Backward-compatibility analysis:

• Existing memory usage sample, list, summary, diff, trend, and leak-signal workflows continue to work.
• memory usage leak-signal keeps the same command name and remains heuristic, but its default text output is now richer because it includes confidence/scope/quality notes.
• The new --since SAMPLE_ID flag is additive and does not change daemon lifecycle or stored sample format.

Risks:

• Scripts that parse the old compact memory usage leak-signal text output may need updates because the default formatter now prints confidence and quality-note text.
• The heuristic is still based on coarse heap-usage checkpoints, so workloads without a clear baseline/action/GC sequence can still produce low-confidence noise.
• Because bounded leak analysis is opt-in, users who keep using full-history mode may still see mixed-session warnings until they adopt --since.

Manual testing:

• Setup: start the daemon, select a target, and leave existing sample history in place to confirm mixed-history messaging still appears.
• Run agent-cdp memory usage sample --label baseline --gc, trigger the suspicious interaction, run agent-cdp memory usage sample --label action, then run agent-cdp memory usage sample --label cleanup --gc.
• Run agent-cdp memory usage leak-signal and verify the output calls out full-history scope or other low-confidence quality notes when applicable.
• Run agent-cdp memory usage leak-signal --since <baseline-sample-id> and verify the output reports bounded scope, includes confidence, and explains the post-GC retained-growth evidence for just that interaction window.
• Expected user-visible outcome: bounded leak checks read as more trustworthy, while weak or mixed evidence is explicitly labeled as such.

Fixes #17

[V3RON]

Manual verification against the PR branch on iOS simulator looks good.

Setup

• built agent-cdp from this PR
• restarted the daemon from the PR build
• launched the React Native playground app in the iPhone 17 Pro simulator
• drove the new Memory Leak Test screen with agent-device

Scenarios run

1. Mixed history + temporary allocation

Preloaded daemon history with unrelated samples, then ran a bounded temporary-allocation workflow.

Commands:

pnpm agent-cdp memory usage sample --label old-1
pnpm agent-cdp memory usage sample --label old-2 --gc
pnpm agent-cdp memory usage sample --label temp-baseline --gc
pnpm agent-cdp memory usage sample --label temp-action
pnpm agent-cdp memory usage sample --label temp-cleanup --gc
pnpm agent-cdp memory usage leak-signal
pnpm agent-cdp memory usage leak-signal --since jm_3

Observed:

NONE confidence:medium score:0 scope:full-history samples:5
note: This result spans all stored samples in the daemon. Mixed workflows can skew the signal; rerun with --since SAMPLE_ID for one bounded check.

NONE confidence:medium score:0 scope:bounded samples:3
note: Only 3 samples in this window; leak confidence is limited.

This matches the intended behavior: full-history mode warns about mixed samples, while the bounded temporary-allocation window stays negative.

2. Retained allocation positive case

Used the Allocate Retained 20 MB x5 action to force obvious retained growth.

Commands:

pnpm agent-cdp memory usage sample --label retained-baseline --gc
pnpm agent-cdp memory usage sample --label retained-action
pnpm agent-cdp memory usage sample --label retained-cleanup --gc
pnpm agent-cdp memory usage leak-signal --since jm_6
pnpm agent-cdp memory usage leak-signal --since jm_6 --verbose

Observed compact output:

HIGH confidence:medium score:5 scope:bounded samples:3
note: Only 3 samples in this window; leak confidence is limited.

Observed verbose output:

JS Memory Leak Signal:
  Level:  HIGH (score: 5)
  Confidence: MEDIUM
  Scope: bounded window (3 samples, jm_6 -> jm_8)

Evidence:
  - Window jm_6 -> jm_8: baseline 6.7 MB, peak 108.0 MB, latest 107.0 MB.
  - Post-GC checkpoint jm_8 is +100.3 MB vs baseline after +0.9 MB of recovery from the peak.
  - Most peak growth remains after a GC-assisted checkpoint, which is a strong retention signal.
  - Latest heap is still 1500.0% above the baseline.

Quality notes:
  - Only 3 samples in this window; leak confidence is limited.
  - Samples oscillate instead of following a clean progression, which lowers confidence.

Caveat: This is a heuristic signal based on heap usage checkpoints, not proof of a memory leak. Use heap snapshots for confirmation.

This is the strongest confirmation for the PR: the positive result is driven by the post-GC retained-growth checkpoint, and the output explicitly communicates confidence, scope, quality notes, and caveat text.

3. Trend-only bounded run without a GC checkpoint

Ran a bounded two-sample retained-growth window without any GC-assisted sample in the window.

Commands:

pnpm agent-cdp memory usage sample --label trend-baseline
pnpm agent-cdp memory usage sample --label trend-action
pnpm agent-cdp memory usage leak-signal --since jm_11
pnpm agent-cdp memory usage leak-signal --since jm_11 --verbose

Observed:

MEDIUM confidence:low score:3 scope:bounded samples:2
note: Only 2 samples in this window; leak confidence is limited.

Verbose quality note correctly called out the missing GC checkpoint:

- No GC-assisted checkpoint in this window; the signal is trend-based and lower confidence.

This also matches the PR intent.

4. Too few samples

Command:

pnpm agent-cdp memory usage leak-signal --since jm_12

Observed:

NONE confidence:low score:0 scope:bounded samples:1
note: Capture a bounded baseline and follow-up sample before using leak-signal.

5. Invalid sample id

Command:

pnpm agent-cdp memory usage leak-signal --since does-not-exist

Observed:

Sample does-not-exist not found

Conclusion

Manual device testing supports the change set:

• --since SAMPLE_ID correctly scopes analysis to one bounded interaction window
• full-history mode warns about mixed daemon history
• retained post-GC growth drives strong signals
• trend-only and too-few-sample windows are explicitly lower-confidence
• invalid --since fails clearly

One minor observation: the mixed full-history temporary case reported confidence:medium rather than low, even though it did include the correct mixed-history warning. Worth double-checking whether that confidence level is exactly what we want, but the core behavior looks correct.