v0.4.5

Added

• TELA on-chain tooling — two new tools take TELA from docs-only to a
  read-only builder surface. tela_inspect fetches a SCID and parses it as a
  TELA-INDEX-1 app manifest or a TELA-DOC-1 file contract (auto-detected),
  enumerating the full ordered DOC1..DOCn list, mods, commit/version history,
  and an honest immutability note; non-TELA SCIDs return kind: not_tela
  (a success, not an error). tela_get_doc_content extracts a DOC's actual
  file content from the contract's DVM comment block, with offset pagination
  and .gz-compression flagging. Both read the raw DERO.GetSC stringkeys
  directly via a new shared parser (src/tela-parse.ts) so a large manifest's
  DOC list is not truncated by the 50-key surface cap. Total tools: 28 → 30.

Changed

• explain_smart_contract is now TELA-aware. TELA contracts are detected
  before the token/registry heuristics, fixing a misclassification where a
  TELA contract's EXISTS("nameHdr") made it register as a name registry. The
  kind union gains tela_index and tela_doc, and TELA contracts now cite
  the TELA spec pages.
• Docs search rewritten from substring matching to a BM25F scorer. The old
  binary .includes() scorer (no tokenization, stopwords, term frequency, or
  length normalization) ranked the wrong pages for real questions — e.g.
  "dero vs monero" matched the word "vs" across 144 of 147 pages. The new
  scorer adds word-boundary tokenization (incl. hyphenated standard names like
  TELA-INDEX-1), a curated stoplist, per-field length normalization (which
  ends the Captain-archive keyword-sponge problem), and IDF weighting so rare
  discriminating terms steer ranking. The previously-unused description field
  is now scored, and excerpts are query-centered. recommend_docs_path gains a
  narrow beginner-intent nudge that surfaces the "Understanding DERO"
  orientation page. Computed in-process at load time — no index-format change.

Fixed

• explain_smart_contract / dero_get_sc no longer overflow host token
  limits on large registries. The name service's 22,619 stored keys produced
  a ~413 KB response (rejected by MCP hosts) on the very SCID the tool docs
  recommend as the known-good example. State-variable maps are now capped at a
  sampled 50 keys with *_total / *_truncated markers.
• Tool failures now set isError: true at the protocol level, so MCP hosts
  that branch on the flag recognize a failed call (the structured
  ok:false/_meta.error body is unchanged).
• Two MCP prompt arguments could never validate. reference_topoheight
  (z.number()) and include_breakdown (z.boolean()) were declared with
  non-string zod types, but prompt arguments arrive as strings; they now coerce
  correctly.
• Metadata drift across the front door. README version, tool/resource/prompt
  counts, Node floor, and the registry description were stale; all corrected and
  the surface counts are now guarded by check:server-json.

Distribution

• The live mcp.derod.org streamable-HTTP endpoint is now published in the
  registry via a remotes entry in server.json.

Tests

• New offline guards wired into CI: check:tela-parse (21 TELA-parser fixture
  assertions, incl. a >50-DOC INDEX proving cap-bypass and a registry-is-not-TELA
  case) and check:docs-ranking (the 6 confirmed search cases as a before/after
  regression harness). A live tela_inspect false-positive guard asserts the
  22,619-key name registry classifies as not_tela.