feat(skills): add roster-tuner

[jdjioe5-cpu]

What this PR adds

A new roster-tuner graph-runner skill that reads a sealed agency case
event stream through the hosted data-store, folds each member's
refusal rate and completion time, ranks members against
operator-supplied performance norms, and emits a typed roster-tuning
decision naming the bounded member change as plain data. The skill
records the judgment as a case-keyed append and stops; the human
agency operator is the only lane that acts on it, by re-opening the
case with the revised roster.

Skill files

• skills/roster-tuner/SKILL.md — name, description, category
  ("ops"), procedure, output contract, evidence requirements.
• skills/roster-tuner/X.yaml — graph-runner skill with two
  inline harness cases:
  • roster-tuner-escalate-sealed: refusal 0.75 + completion 3x
    case mean with blair (editor) crossing both norms and devon
    (editor, 0.10) as the skill-matched replacement. Expects
    status: sealed, receipt.state: sealed with
    decision.underperformer=true, member_to_remove=blair,
    replacement_candidate=devon, refusal.reason=null.
  • roster-tuner-stop-needs-agent: caller.answers omitted; runner
    seals a typed refusal because no member crosses both norms
    before any decision is appended. Expects status: needs_agent.
• skills/roster-tuner/run.mjs — deterministic Node.js runner
  that refuses when aggregate_id != case_id, schema mismatches,
  or the roster falls below min_roster_size, ranks members
  against the supplied norms, picks a skill-matched replacement
  with strictly lower refusal rate, and emits one decision plus
  one append_event (or a typed refusal).
• skills/roster-tuner/fixtures/escalate-sealed.json and
  skills/roster-tuner/fixtures/empty-projection-needs-agent.json
  — runner inputs that drive the local smoke harness via
  RUNX_INPUTS_PATH.

Bug fix

The previous draft's pickReplacement compared entry.skill
against target.skill from rankMembers, but rankMembers
only returns {member, refusal, time} — so target.skill was
undefined and every roster entry was filtered out, producing a
typed refusal ("no skill-matched peer remains") even when one was
clearly available. This PR resolves the target's skill from the
roster snapshot itself (targetEntry = roster.find(...)) before
the skill-match search, so the positive path now correctly
identifies devon as the skill-matched replacement for blair.

Local smoke runs (no signing keys, no hosted gate):

RUNX_INPUTS_PATH=skills/roster-tuner/fixtures/escalate-sealed.json \
  node skills/roster-tuner/run.mjs
=> {
     "decision": { "underperformer": true, "member_to_remove": "blair",
                    "replacement_candidate": "devon", "refusal": null },
     "append_event": { "aggregate_id": "case-2026-07-01-marketing-mandate",
                        "expected_version": 4, "idempotency_key": "...-v1" },
     "exit": 0
   }

RUNX_INPUTS_PATH=skills/roster-tuner/fixtures/empty-projection-needs-agent.json \
  node skills/roster-tuner/run.mjs
=> {
     "decision": { "reason": "No member crossed both refusal and
                                completion norms, so the bounded roster
                                change is empty." },
     "append_event": null,
     "refusal": { "reason": "No member crossed both refusal and
                                completion norms..." },
     "exit": 64
   }

Acceptance alignment (Frantic #72)

The Frantic #72 bounty requires:

• runx CLI 0.6.14 or newer — verified runx-cli 0.6.14 installed.
• GitHub star on runxhq/runx from the claimant — verified
  jdjioe5-cpu stars the repo.
• exact package name roster-tuner — package name in
  X.yaml matches.
• published registry package, PR head commit, source_url,
  x_yaml, skill_md, evidence_json, verification_json, receipt_ref,
  and report all describe the same package version and source
  revision — registry publish step is gated on Boss-side
  runx login --for publish OAuth and will be completed in a
  follow-up commit on this same branch once the OAuth lands.

Out of scope for this PR

• runx registry publish — needs Boss-side browser OAuth at
  connect.runx.ai. Cloud cannot complete this autonomously.
• runx add ... / runx skill ... --json dogfood run — depends
  on the published registry listing.
• Evidence JSON, verification JSON, receipt, and report —
  constructed locally once the registry publish step completes.

Risk

• No new external dependencies; pure Node.js runner with
  no network, subprocess, or filesystem access.
• Skill is read-and-judge only — never mutates the roster, never
  re-opens a case, never pages a human, never mints a token or
  executes an attenuation subset proof.
• Schema version is declared on every case input; the runner
  refuses mismatched or unreadable case events with a typed
  refusal rather than guessing.