feat(pr-workflow): add /mms-debug orchestrator

domains/pr-workflow/skills/debug/skill.md

@@ -0,0 +1,101 @@
+---
+name: debug
+description: Locate the cause of a symptom you cannot yet explain — a crash, a leak, a flake, a production error spike, a number that moved. The sibling of evidence: where evidence is handed a claim and looks for the observation that would falsify it, this is handed a symptom and must generate the hypothesis first, then kill it. Classifies the symptom into a defect class, routes to the engine skill that owns that class (memory-leak, race-condition-repro, react-render-proof, sentry-grafana-correlation, extension-errors-debugging, tsc-blindspots, supply-chain-audit), and holds the investigation to the same evidence bar evidence applies — an instrument that cannot fail is not evidence, a null needs its sensitivity stated, and a finding is scoped to what the change introduced versus what pre-existed. Stops when the cause is located or the class is excluded, not when a plausible story is available. Triggers on /mms-debug, or when asked to debug, diagnose, or investigate a symptom, find why something is slow, leaking, flaky, or erroring, chase a production alert to its cause, or reproduce a bug that cannot be reproduced by hand.
+maturity: experimental
+---
+
+# /debug
+
+`evidence` is given a claim and looks for the observation that would prove it false.
+This is given a **symptom** and has to produce the hypothesis before anything can falsify it.
+
+That difference is the whole skill. In review, the claim is someone else's and the social
+pressure runs toward scepticism. In debugging, the hypothesis is *yours*, nobody else is
+positioned to challenge it, and the expensive failure is building three hours of work on the
+first theory that fit the first observation.
+
+## When To Use
+
+- A symptom with no established cause: a crash, a hang, a leak, an intermittent test, an
+  error-rate step change, a metric that moved without a deploy that explains it.
+- A bug you cannot reproduce by hand and therefore cannot yet observe.
+- A production signal that needs chasing back to code.
+
+## Do Not Use When
+
+- The PR states a claim and you need it settled — that is `/evidence`.
+- The cause is known and you are validating the fix — that is `/evidence`, or the engine
+  skill directly.
+- You want an after-the-fact writeup of a resolved failure — that is a postmortem, not this.
+
+## Workflow
+
+1. **State the symptom as an observation, not a theory.** "Popup memory grows ~105 MB per
+   open/close cycle" — not "the popup leaks because of the snow hook". The theory is the
+   output, never the input.
+2. **Classify into a defect class** (table below). If two classes fit, run both; do not pick
+   the one you find more interesting.
+3. **Delegate to the engine.** Each owns its own method and its own falsifier. This skill
+   routes and holds the bar; it does not re-implement the investigation.
+4. **Kill the hypothesis before extending it.** Name the observation that would rule it out,
+   and go looking for that observation specifically. A hypothesis that has only ever been
+   confirmed has not been tested.
+5. **Stop on a located cause or an excluded class.** A plausible story is not a stop condition.
+
+## Symptom → engine
+
+| Symptom | Class | Engine |
+|---|---|---|
+| Memory grows across a repeated flow; tab or worker dies over time | retention | `memory-leak` |
+| Intermittent failure; passes on rerun; order-dependent | interleaving | `race-condition-repro` |
+| UI janks, re-renders excessively, selector recomputes | wasted render work | `react-render-proof` |
+| Production error spike, latency change, or a metric that moved | production signal | `sentry-grafana-correlation` |
+| Extension-specific: MV3 vs MV2, background vs UI context, service-worker lifecycle | platform | `extension-errors-debugging` |
+| Runtime value disagrees with its declared type; green typecheck, wrong behaviour | type/reality drift | `tsc-blindspots` |
+| Started after a dependency change; new capability or transitive edge | supply chain | `supply-chain-audit` |
+| None of the above, or several | — | bisect to a change first, then re-classify |
+
+## The evidence bar carries over
+
+The engines are shared with `evidence`, and so are its trust gates. They matter more here,
+because in review a weak instrument produces a weak claim someone else will challenge — in
+debugging it produces a wrong theory nobody checks.
+
+- **An instrument that cannot fail is not evidence.** Before trusting a measurement, establish
+  it can report the negative: a positive control that must move, a base arm that must fail.
+- **A null needs its sensitivity stated.** "No difference" and "could not have detected one"
+  print identically. Calibrate, then report the zero against what the instrument demonstrably
+  resolves.
+- **Scope to the change.** Classify each finding as introduced-here versus pre-existing.
+  Report pre-existing separately and uncharged, or you will attribute an old defect to a new
+  diff.
+- **A negative result carries the scope of its search.** "No leak found", "nothing in the logs",
+  "the artifact does not exist" are claims about where you looked. Name the stores searched in the
+  finding itself — filesystem, artifact bucket, issue tracker, the other process's logs. If you
+  cannot name them, the search is not finished.
+- **Measure on an isolated machine.** Timing- and GC-sensitive numbers taken on a contended host
+  are not noisy-but-usable, they are *stably wrong* — several runs will agree with each other and
+  disagree with reality. Replicate across hosts, not just across runs, before trusting a figure.
+- **Collect the whole battery, not the discriminating member.** Instruments that stay flat are
+  data: the joint pattern localises the defect in a way no single reading does.
+
+## Output
+
+A short investigation record, not a narrative:
+
+```
+SYMPTOM    observation, as measured
+CLASS      defect class, and why (with the classes considered and dropped)
+HYPOTHESES each with the observation that would kill it, and whether that was found
+CAUSE      located mechanism, at file:line — or "class excluded", which is a real result
+NOT CAUSE  hypotheses killed, kept so the next person does not re-walk them
+```
+
+Killed hypotheses are part of the deliverable. Deleting them makes the surviving one look
+inevitable and hands the next investigator the same dead ends.
+
+## Scope — what this is NOT
+
+- Not a fix. It locates and evidences the cause; the change is a separate act.
+- Not a replacement for the engines. Each owns its method; this chooses among them.
+- Not an incident-management process. No severity, comms, or timeline.