v0.8.7

[0.8.7] — 2026-05-19

Fixed

• Parsed-DataFrame cache below the byte cache — 12x speedup on warm
  F-table reads (Bug 7, 2026-05-19 latency diagnostic). _get_data_impl
  previously called pd.read_csv on the cached CSV bytes for every
  invocation. The byte cache hit in <10 ms but parsing the 300 KB F1
  CSV with ragged-row handling + numeric coercion took 5+ seconds on a
  shared CPU. Net effect: rba.F1.latest() p50 = 5.2 s warm, which
  set the floor for the ausdata-api /v1/real-cash-rate p50 (4.75 s
  warm) and bumped p99 to 6.5 s.

  Added an LRU-bounded (csv_filename, md5(body)) -> (header, DataFrame)
  cache layer between the byte cache and the parser. Bounded at 10
  entries (~300 MB worst case at 30 MB/parsed-frame). Cache is
  module-global (shared across worker threads) because every thread
  parses the same upstream bytes to the same DataFrame — sharing the
  parse is a strict win on the gateway hot-path. Fresh upstream
  publications miss cleanly via the md5 key.

  Also routes describe_table() through the shared cache so describing
  a recently-fetched table is ~10 ms instead of seconds.

  New reset_df_cache_for_tests() helper for clean test isolation
  (mirrors asic-mcp's parquet-cache reset pattern). New
  tests/test_df_cache_hits_on_same_csv.py pins the cache behaviour:
  N identical-bytes calls trigger exactly 1 parse; differing bytes
  re-parse; LRU caps at _DF_CACHE_MAX_ENTRIES; reset clears state.

  Expected impact at ausdata-api: /v1/real-cash-rate p50 drops from
  4.75 s -> <500 ms; p99 from 6.5 s -> <1.5 s.

0.8.5 — F1 (Daily Cash Rate) curated

Customer-sim flagged that rba.F1.1.latest() returns the month-end snapshot (4.10%) instead of the current cash rate (4.35%, set 2026-05-06). Added F1 as curated — F1.latest() returns the freshest daily reading. 184 unit tests pass.

rba-mcp 0.1.9 — describe_table freshness + yield-curve search routing

0.1.9 (2026-05-13)

Loop-audit value pass — three low-effort, high-value polish wins surfaced
by a focused review of the customer-facing surface.

• New: SeriesDetail.end_date — describe_table now reports the
  latest non-null observation per series alongside start_date. An
  LLM can answer "is this data fresh?" from a single describe_table
  call without a follow-up latest(). Populated via
  df[col].last_valid_index() in both curated and non-curated branches.
• Search: F2 / F2.1 keywords expanded. Added yield curve,
  bond yields, australian government securities, ags, and short
  bond tenors (2 year bond, 3 year bond, 5 year bond) to the
  f2-data and f2.1-data entries in tables.yaml. A user asking
  "what's the Australian yield curve?" now lands on F2 / F2.1
  immediately. Verified via two new search routing tests.
• Search ranker: phrase-match bonus (+20). Added alongside the
  existing curated +30 bonus. If the full query phrase appears as a
  substring in a table's haystack, that table gets +20. Lets strong
  non-curated matches (like "yield curve" → F2) compete with the
  curated boost when a query is highly specific to a non-curated
  table; curated tables that ALSO phrase-match still stack both
  bonuses, so common queries route correctly.
• Tests: +4 regressions — end_date populated for curated +
  non-curated, "yield curve" → F2 routing, "bond yields" → F2
  routing. 111 unit tests now (was 107).

rba-mcp 0.1.8 — customer-flow audit fixes

0.1.8 (2026-05-12)

Customer-flow audit fixes — surfaced when running rba-mcp against Claude
Desktop end-to-end. Two real UX gaps + a major distribution-side warning.

• Fix: start_date and end_date accept int years. MCP / LLM clients
  often send a year as a JSON number (start_date=2024) instead of a
  string ("2024"). Pre-0.1.8 this errored at the Pydantic boundary
  with a verbose "Input should be a valid string" message. Now: the
  Annotated type is str | int | None and _validate_period coerces
  int → str transparently. Bool is explicitly excluded from coercion
  (since isinstance(True, int) is True in Python) so True/False
  still raise a clean type error rather than becoming "1"/"0".
• Fix: describe_table populates SeriesDetail.start_date for
  curated tables. The non-curated branch already fetched the CSV and
  ran df[sid].first_valid_index(); the curated branch shortcut to
  YAML metadata only and left start_date null on every curated
  series. The LLM needs this to pick a sensible date range without
  trial-and-error queries — e.g. F11 starts 2010, F11.1 starts 2023.
  Now both branches populate the field.
• README + example configs: recommend uvx --upgrade rba-mcp for
  Claude Desktop / Cursor. Plain uvx rba-mcp (no flag) uses
  whatever wheel is cached and the long-lived MCP child process holds
  it — new PyPI releases never reach a running install until the user
  manually uvx --refreshes and fully quits Claude Desktop. Found in
  the wild during a customer-flow audit: an install was running 0.1.2
  against PyPI 0.1.7, five releases of fixes never adopted. --upgrade
  makes uvx check PyPI on each launch. Same fix shipped to abs-mcp's
  example configs.
• Tests: +4 regressions in tests/test_stress_regressions.py
  (int year accepted; int = str equivalence; bool still rejected;
  curated start_date populated). 107 unit tests now (was 103).