refactor: dedupe the opencode DB preflight availability check

[sorafujitani]

Context

The decision "is the OpenCode DB usable, and if not, is it a hard error or just absent?" (including the legacy-storage detection and the OPENCODE_DB override rule) exists in two places in internal/source and can drift.

Current state

• internal/source/opencode.go:27-44 — Preflight for --source=opencode
• internal/source/all.go:37-53 — newOptionalOpencodeSource for --source=all

Both implement the same shape: on ErrDBNotFound with OPENCODE_DB unset, check for legacy storage and branch.

Task

1. Read both functions and write down the exact decision table (error kind × env set × legacy present → outcome). The two callers want different outputs (hard error with guidance vs. silently-skip), so extract the classification, not the reaction.
2. Add a single classifier to internal/opencode, e.g.:

   type Availability int
   const (
       Available Availability = iota
       NotFound          // no DB, no legacy storage
       LegacyOnly        // legacy storage present, new DB absent
       Misconfigured     // OPENCODE_DB set but unusable
   )
   func ClassifyAvailability() (Availability, error)

   (Adjust names/cases to match the real decision table from step 1.)
3. Rewrite both call sites as thin switches over the classification; keep the user-facing error messages byte-identical.
4. Add a table-driven test for the classifier using temp dirs / env overrides (OPENCODE_DB, fake legacy layout). Check existing tests in internal/source for these paths and de-duplicate.

Acceptance criteria

• The legacy-detection logic exists in exactly one place.
• Error messages for --source=opencode without a DB are unchanged (test-asserted).
• go test ./..., go vet ./..., golangci-lint run pass.