Add GVL/allocation instrumentation and optional log fields

[Peralphix]

Extracted GVL/memory instrumentation from our applications into the shared gem.

Added new files (loaded only via explicit require):

• patches/active_support_event — adds #gvl_time and #malloc_increase_bytes to ActiveSupport::Notifications::Event;
• patches/rails_semantic_logger — enriches the "Completed" action log entry with GC/GVL/allocation stats;
• semantic_logger/sidekiq_job_metrics — same stats for Sidekiq jobs.

TinyJsonFormatter now supports optional duration and payload keys in custom_names_mapping. Previously such keys were silently ignored. Output is unchanged unless the keys are explicitly requested.

[tycooon]

🤖 AI-assisted review (Claude Code), non-blocking.

Nice that the patches are opt-in (require-your-own) and covered by specs. Two small notes:

1. GC.stat[:malloc_increase_bytes] allocates the whole stat hash. In lib/umbrellio_utils/patches/active_support_event.rb, start!/finish! call GC.stat[:malloc_increase_bytes], which builds the entire GC.stat hash (~30 entries) each time — twice per instrumented event. GC.stat(:malloc_increase_bytes) reads just that one key without allocating the hash. A little ironic for allocation instrumentation, and cheap to change.

2. malloc_increase_bytes can be negative / cross-thread. It's a delta of a process-global GC counter, so it (a) can drop below zero if a GC runs inside the window, and (b) under threaded servers (Puma threads, Sidekiq concurrency) includes allocations from sibling threads. The consuming yabeda PRs guard with .positive?, but Event#malloc_increase_bytes surfacing negatives might deserve a doc caveat. gvl_time (GVLTools LocalTimer) is per-thread, so that one is fine.

[tycooon-review-bot]

Claude Code (AI review) — reviewed against origin/main.

Clean, well-scoped extraction. Verified the parts that are easy to get wrong: the three instrumentation files are genuinely opt-in (the gem's entrypoint uses explicit require_relative and doesn't load them, so require "gvltools" only fires for consumers who opt in), GVLTools::LocalTimer.monotonic_time is nanoseconds so / 1_000_000.0 → ms is correct, GC.stat(:malloc_increase_bytes) reads the single key without allocating the stat hash, and the base gem really does define extract_path (the spec stub mirrors it). The prior review's two notes (GC.stat allocation, negative/cross-thread malloc) are already addressed in the current code.

Two non-blocking notes:

• P1 — add_stats (rails_semantic_logger) and metrics_payload (sidekiq_job_metrics) duplicate the same 6-stat extraction block plus a second PRECISION = 6; worth extracting a shared helper.
• P2 — the controller patch leaves the raw :exception_object in the logged payload on failed actions, inconsistent with the Sidekiq path (which keeps only [class, message]).

No P0s. The TinyJsonFormatter optional-fields change is backward-compatible. CI is green across Ruby 3.3 / 3.4 / 4.0.

[tycooon-review-bot]

APPROVED — Claude Code (AI review). Both findings addressed in 25fcc21 with tests (P1: shared ActiveSupportEvent#stats; P2: :exception_object dropped + failed-action spec). CI green on Ruby 3.3/3.4/4.0.