fix: auto-resolve two bug reports — plugin reattach isolation (#496) + routed-model usage attribution (#495)

[TYRMars]

Auto-resolves two independent bug reports, one commit each.

---

#496 — plugin reattach: one bad SKILL.md silently kills a plugin's MCP servers + sibling skills

PluginManager.#reattachOne iterated a plugin's skills with no per-skill error guard, and the MCP-server loop ran after it. On server restart, one unreadable/unparseable SKILL.md threw and aborted the rest of that plugin's reattach — silently killing every MCP server it declared plus all sibling skills — while reattachInstalled's bare .catch(() => {}) swallowed the error and GET /v1/plugins kept reporting the plugin as fully installed (that data comes from the ledger, not live state).

Fix (packages/plugin/src/manager.ts):

• Per-skill try/catch around each skill's load/parse/insert, so one bad file can't take down siblings or MCP servers (matches SkillCatalog.scanRoot and the MCP loop's existing best-effort pattern).
• MCP add failures recorded rather than silently swallowed.
• New degraded / last_error fields on the ledger entry, set at reattach time and cleared on a clean reattach, surfaced through list() / get() so GET /v1/plugins stops reporting a half-attached plugin as healthy.
• reattachInstalled records whole-plugin failures (e.g. corrupt plugin.json) on the entry.
• Web-side InstalledPlugin type gains the two optional fields.

Tests: a corrupted installed SKILL.md is skipped while its sibling skill and the MCP server still come up and the entry is marked degraded; a subsequent clean reattach clears the marker.

---

#495 — routed usage priced against the wrong model under JARVIS_ROUTER_ENABLED

The agent loop stamped every usage event with the static config.model, not the model the request actually ran on. With routing enabled, RoutingProvider rewrites the model before delegating, so all token usage from routed turns was attributed to the pre-routing model and the web UsagePanel priced it from a table keyed by that wrong name — under-reporting cost by up to two orders of magnitude, invisibly (the expensive tier showed zero usage; the "unknown model → partial estimate" safety valve never fired because the stale name is a valid known one).

Fix (packages/core, packages/router):

• Optional model added to the usage LlmChunk variant and to ChatResponse (absent for plain providers).
• RoutingProvider tags the streamed usage chunk and the blocking response with the resolved tier target, using ?? existing so a downstream that reports its own model wins.
• agent.ts prefers chunk.model ?? this.config.model — non-routing providers unchanged, routed usage attributed correctly.

Tests: streamed + blocking routed usage carries the tier-target model; a downstream-set model is not overwritten.

---

Verification

• pnpm --filter @jarvis/plugin test → 30/30; @jarvis/router → 27/27; @jarvis/core → 30/30. Lint clean on all changed files.
• Pre-existing, unrelated failures in this environment are not touched: server tests needing native pty.node / an unbuilt UI dist, and a latent assert.ok type-narrowing error in router.test.ts's chain-reconciliation test (present on main, identical before/after this change).

🤖 Generated with Claude Code

https://claude.ai/code/session_01Pdj9CWcTPhxPMi559vJpDz