feat(servermonitor): dedupe sub-process report, add cursor + in-memory merge

[szhatchenko]

Summary

The "Slow Sub-Processes (during profile window)" section of the profile summary
previously printed one line per scan, so a process alive across many scan ticks was
repeated dozens of times, a process that finished between two scans was missed
entirely, and a process that outlived the profile window was not attributable to that
profile at all.

This PR makes the section per-pid and cumulative:

1. Deduplicate per pid across scans. Each entry now carries first/last observation
   time plus an estimated lifetime (observed interval + age delta) instead of a single
   snapshot age.

2. Cumulative report cursor. A high-water mark (subProcessLogCursor) is persisted
   in the profile's .json metadata sidecar. The first report covers the profile's time
   window; each subsequent report shows only what was observed since the previous
   report, so long-running external processes are still attributed to the profile that
   triggered them. The cursor survives server restarts (worst case: a few lines shown
   twice, harmless because entries are deduplicated per pid).

3. In-memory observation registry. MonitoringService now tracks each pid's
   first/last-seen and max age across scans (pruned after 2 consecutive missed scans, so
   one transient /proc read failure does not drop a process). The report merges this
   live view with the persistent subprocesses.jsonl log, covering processes still
   observable right now — including ones that started after the profile window ended.

Example new report line

2026-08-31 14:02:11 → 2026-08-31 14:26:47  pid=1234 lifetime~1476s (max observed age 1450s)  perl script.pl --opt1

Files

• MonitoringService.java — SubProcessObservation registry, updateSubProcessObservations(), getObservedSubProcesses()
• SubProcessMonitor.java — first/last-seen + age fields, estimatedLifetimeSec()
• SupportServlet.java — appendSubProcessSummary() rewrite (dedup + cursor + memory merge), persistSubProcessCursor(), SubProcessEntry

Verification

• mvn -pl src compile — clean
• ant -f src/build.xml compile — BUILD SUCCESSFUL (the unchecked-operations note on SupportServlet is pre-existing)
• TestSubProcessLog — OK (8 tests)
• TestAsyncProfilerWrapper — OK (5 tests)

Single commit on top of main (replaces the superseded #32, which was mistakenly based on a gxp_dev-side cherry-pick branch and therefore carried 50 unrelated commits).

Co-Authored-By: Claude Code noreply@anthropic.com

[szhatchenko]

Review fixes — addressed in amended commit ae26f269

Thanks for the thorough review. All points addressed:

Blockers

• Exchange some tracks formats #1 (first report pulls in unrelated in-memory processes) — getObservedSubProcesses now filters by interval overlap (lastSeen >= intervalStart && firstSeen <= intervalEnd) instead of an unbounded "since service start" read. The first report uses the same ±120 s effective window for both the log read and the in-memory test. A process observed at 15:00 but still alive at 17:05 now only appears if it overlaps the profile window; one that started before and is gone by the window is excluded.
• Feature/cram track importer #3 (lifetime double-counted) — ageSeconds is now − startInstant (total elapsed since process start), so the total-lifetime estimate is simply lastAgeSec. The old interval + (lastAge−firstAge) double-counted the same wall-clock span. Fixed in estimatedLifetimeSec().
• solve problem wth cram #4 (log entries showed lifetime~-1s) — each log snapshot now passes its own ageSeconds as the lifetime estimate, so a process present only in the persistent log (no longer in memory) still gets a real lifetime (the max observed age), not -1.
• Translate web interface into russian language #2 (cursor semantics contradicted the filter) — resolved by the same overlap fix: a follow-up report uses [cursor, now], and a process alive at the cursor instant overlaps it (lastSeen >= cursor) so it is included, matching the documented intent.

Non-blockers

• add CRAM format #5 (PID-only identity) — documented as an accepted limitation in appendSubProcessSummary / SubProcessEntry / mergeSubProcessEntry Javadoc (PID reuse within the short window can merge two unrelated processes).
• Gxp stable #6 (everSlow semantics) — field renamed in spirit to slow but documented as "exceeded the threshold at least once"; the report header now states everSlow = exceeded the slow threshold at least once.
• Gxp stable #7 (empty-scan interpretation) — verified SubProcessMonitor.check() swallows per-process exceptions and returns what it can, so the grace period (2 consecutive missed scans) tolerates a transient /proc failure; an empty scan is treated as authoritative only for pruning, never for the report.
• perf(profiler): optimize CRClusterAnalysis hot paths from async-profiler #8 (non-atomic cursor write) — persistSubProcessCursor now writes to a temp file in the same directory and atomically renames over the sidecar (with a non-atomic fallback), and works on a copy so the caller's JSONObject is never mutated.
• fix(server): pass oldValue instead of property to correctBeanOptions #9 (cursor naming/semantics) — renamed subProcessLogCursor → subProcessReportCursor and documented it as a report-time cursor (the instant the report was generated), not a log offset.
• perf(profiler): optimize hot paths from async-profiler #10 (O(n) truncation count) — slow count is now computed once before the render loop.

Tests added (TestSubProcessLog, now 12 tests, all passing):

• testEstimatedLifetimeUsesLastAgeOnly — lifetime = last observed age, not interval+ageDelta; single snapshot → -1.
• testObservationOverlapFilter — the four attribution scenarios (started-before-alive-in-window, ended-before-window, started-after-window, spans-window).
• testMergeLifetimeAndCommand — merging log snapshots yields a real lifetime (max age), earliest firstSeen / latest lastSeen, sticky everSlow, kept command.
• testPersistCursorAtomicAndCopy — cursor file written, original meta not mutated, no temp files left, non-advancing cursor is a no-op, missing parent dir leaves no partial file.

Verification: mvn -pl src compile clean; ant -f src/build.xml compile BUILD SUCCESSFUL; TestSubProcessLog OK (12); TestAsyncProfilerWrapper OK (5).