feat(skills): add triage skill and plugin manifest

[h30s]

New skill: triage (Closes #5)

This PR adds the triage skill to help turn logs and crash traces into ranked hypotheses with tests that tell them apart, without jumping to conclusions or assuming correlation is causation.

What this skill does

• Quotes exact log lines before making any inferences.
• Ranks possible causes and defines a clear test to distinguish between them.
• Calls out log limits explicitly (like buffer truncation or missing debug levels).
• Treats correlation as a lead, not a confirmed cause.
• Points to the single next step (specific file, command, or log level) if the log isn't enough.

What task I tested it on and what changed

• Task: Tested against a real crash log that showed an OutOfMemoryError in a background worker right before a container kill (SIGKILL with OOMKilled: false).
• What changed: Without the skill, the immediate guess was to fix an OOM leak in the worker code. With the skill, it quoted the evidence, pointed out the short time window, and separated two competing hypotheses: a long JVM GC pause failing the health check vs. an upstream auth timeout cascading into connection pool exhaustion. It named the distinguishing check (checking GC logs vs. auth response times) instead of making an unverified fix.

What I deliberately left out and why

• Code editing and fixing: Triage stops at finding the ranked hypotheses and distinguishing tests. Actually applying a fix belongs to code editing, not log triage.
• Generic debugging tips: Kept out broad advice like "check system health" and required naming the exact file, command, or log level needed.

[ms-shashank]

I have reviewed this and it is good to merge.

The thing I like most is that the skill knows where it stops. The ECONNREFUSED example ends by admitting the buffer is truncated at the wrong end instead of picking a winner, and "Cannot determine from this log" is written in as an allowed answer. That is the hard part of triage and most attempts at this skip it.

The configuration matches what the skill actually does, which is worth calling out because it is easy to get wrong. tools: chat is correct here since the skill reads and reasons and its final output is naming the decisive command rather than running it. workspace: optional is correct because someone can paste a log with no project open.

One note for merge order. This touches the same lines of .oxcode-plugin/marketplace.json as #11, so whichever goes second needs a rebase. I am merging this one first since it is clean.

Nice work.