feat(skill): add /pramana skill with subcommands over MCP primitives

[sarath-soman]

Motivation

Pramana's primary interaction surface is MCP (from Claude Code) + CLI. MCP tools are already exposed (`list-tenants`, `get`, `search`, `traverse`, `list`), but invoking them mechanically one-at-a-time through the MCP layer misses useful orchestrations: a search that auto-opens the top hit, a "show me everything about X" that runs get + traverse + formatted summary, tenant mounts/reloads, doctor output, etc.

A single `/pramana` skill with dispatch-style subcommands gives one clear entry point, reduces skill sprawl, and keeps the surface discoverable.

Structure

One skill file at `.claude/skills/pramana/SKILL.md` (lives in pramana-ops repo; surfaces in pramana/.claude/skills/pramana/ via the existing symlink). Dispatches on first argument:

```
/pramana [args...]
```

Subcommand	Wraps	Purpose
`tenants`	MCP `list-tenants`	List available KBs with artifact counts
`get [section]`	MCP `get`	Fetch one artifact
`search `	MCP `search`	Rank-ordered search
`list [--tags a,b]`	MCP `list`	List artifacts, optionally filtered
`traverse [--type T] [--depth N]`	MCP `traverse`	Walk relationships
`explore `	get + traverse, formatted	Single-turn deep dive on one artifact
`doctor [--tenant T]`	CLI `pramana doctor`	Integrity diagnostics
`reload `	CLI `pramana reload`	Reload a tenant from disk
`lint [--tenant T]`	CLI `pramana lint`	Lint knowledge sources
`daemon status` / `daemon start` / `daemon stop`	CLI `pramana serve` + HTTP health	Daemon lifecycle
`config`	CLI `pramana config` passthrough	Tenant mounts/edits
`version`	CLI `pramana version`	Version info

Default subcommand (no args or unrecognized) → print a menu of available subcommands and 1-line descriptions.

Skill file layout

```
.claude/skills/pramana/
├── SKILL.md # entry; dispatcher; describes subcommands
├── references/
│ ├── tenants.md # how to read the list-tenants output
│ ├── search.md # ranking semantics, query syntax
│ ├── traverse.md # relationship types, depth safety
│ ├── explore.md # orchestration recipe for the composite subcommand
│ ├── doctor.md # interpreting doctor JSON
│ ├── daemon.md # start/stop, health endpoints
│ └── config.md # mounting tenants
└── commands/ # optional: per-subcommand prompts if they grow
```

Behaviour rules

• Every subcommand that reads from a tenant takes `` as its first positional after the verb, mirroring CLI shape.
• When MCP is available in the session, prefer it over shelling out. Fall back to `pramana ` via Bash only for surfaces MCP doesn't expose (doctor, reload, lint, daemon, config).
• `/pramana` with no args: print the subcommand menu; don't guess intent.
• Don't silently mutate state. `config`, `reload`, `daemon start/stop` should confirm before acting.
• Output formatting: render MCP JSON as human-readable markdown tables/sections in the conversation. Raw JSON only on explicit `--json` or `--raw`.

Out of scope (separate skills / tickets)

• TUI revival or tui-snapshot integration (TUI is being retired — see chore: retire pramana-tui; move to archive/ and cut from CLI + workspace #242)
• Authoring flows (create/edit artifacts) — separate skill
• Cross-tenant federated search — separate ticket, needs engine work first

Acceptance

• `/pramana` dispatch works for every subcommand listed above
• Each subcommand has a one-page reference under `references/`
• Invoking `/pramana tenants` in a fresh Claude conversation returns the live tenant list via MCP (no shell)
• `/pramana explore ` returns a single formatted summary that includes the artifact body, its outbound relationships, and its inverse-relationships, without requiring the user to chain calls manually
• SKILL.md is short (< 200 lines); deep content lives in `references/`