Both surfaces delegate to the single @opencodehub/analysis.runChangePack core,
so they cannot disagree on values.

- MCP change_pack tool: withStore + withNextSteps + staleness + error envelope,
  registered in server.ts; structuredContent recases top-level keys plus the
  interiors of impacted_subgraph and cost_attribution to snake_case.
- CLI codehub change-pack: commander registration, --base/--head/--depth/
  --min-confidence/--budget/--include-tests-in-subgraph/--json; --json is a pure
  passthrough of the raw ChangePack; exit code = verdict.exitCode so CI gate
  semantics match codehub verdict.
- First CLI<->MCP parity test in the repo, split to respect package rootDir:
  the CLI test asserts --json deep-equals the raw pack (passthrough); the MCP
  test asserts toStructured recases losslessly (recase-back == raw pack), so
  together they prove both surfaces serialize identical values.
- server roster contract bumped 28 -> 29 (change_pack added).

The ERPAVal spec + derived task list driving the change_pack feature: the four
output sections, CLI/MCP parity AC, determinism/byte-identity, affected-test
selection semantics, and the char-heuristic cost-attribution decision (tokenizer
is provenance-only in the repo, so cost figures are explicitly estimates).

Two durable ERPAVal lessons from the change-pack session + INDEX pointers:
- tokenizerId is provenance, not an encoder (cost features need a labeled
  heuristic or a real tokenizer dep; the repo ships neither today).
- CLI<->MCP parity via one shared analysis core + a split parity test that
  respects the mcp package rootDir; the server roster contract bumps by design.

…ibution

Replaces the char-heuristic estimate with real BPE token counts via
`gpt-tokenizer` (pure-JS, MIT, zero-dep). Chose it over the native/WASM
`tiktoken` package to honor the no-native-binding rail (ADR 0015); the
`gpt-tokenizer/encoding/o200k_base` subpath bundles its BPE ranks inline, so
`encode` is synchronous and deterministic — byte-identity and the content hash
hold.

- `countTokens` encodes via o200k_base, falling back to `max(1, ceil(len/4))`
  only on input that throws (rare, still deterministic) so cost never crashes
  the pack.
- `CostAttribution` now reports `estimate: false` /
  `tokenizerModel: "openai/o200k_base"`; the type widens `estimate` to boolean
  and `tokenizerModel` to string. The field is folded into the hash, so the
  basis change is visible in `changePackHash`.
- CLI + MCP summaries drop the "(est.)" hedge and name the tokenizer.
- Tests, fixtures, EARS spec U6, and the tokenizer lesson updated to the real
  encoder; license allowlist + OSV clean (gpt-tokenizer is MIT, no advisories).

The CLI bundles `@opencodehub/*` source (noExternal) but keeps third-party
deps external, resolving them from the CLI's own node_modules at runtime. The
change-pack o200k counter added `gpt-tokenizer` to @opencodehub/analysis but
not to the CLI, so `node dist/index.js analyze` threw "Cannot find package
'gpt-tokenizer'" in CI (the self-scan + every fresh-install matrix leg) while
passing locally via the pnpm workspace hoist. Re-declare it on the CLI like the
other analysis third-party deps (@iarna/toml etc.). Per the
tsup-collapse-monorepo lesson: verify from the built dist, not the hot
workspace.

…stic

The cost-attribution tests key an in-memory file map with POSIX paths, but
runChangePack joins repoPath + relPath with the OS separator — so on Windows
the lookup arrived backslash-separated, missed the fixture, and the blind
baseline computed as 0 (two failing assertions on windows-latest). Normalize
the seam's lookup key to forward slashes so the fake reader matches on every
platform. Production fs reads are unaffected (real paths use the right
separator); this is a test-fixture fix only.