v0.13.2

Portfolio Construction. Nothing is removed or renamed: every subpath published
in 0.13.1 resolves unchanged, and no entry point changed its name, signature
or behaviour.

	0.13.1	0.13.2
Topics	697	717
Domains	18	19
Topics with a validated api contract	697 / 697	717 / 717
Topics with an executed worked example	697	717
Topics whose arithmetic is asserted on every build	623	623

New

• Portfolio Construction — 20 topics, the first domain here that turns a
  covariance matrix into a set of weights. Mean-risk optimization (Markowitz
  against a return target, global minimum variance, maximum Sharpe, mean-CVaR,
  mean-MAD), risk allocation (inverse volatility, equal risk contribution, risk
  budgeting, hierarchical risk parity and its equal-risk-contribution variant),
  Bayesian and robust allocation (Black-Litterman, resampled frontier, robust
  and distributionally robust mean-variance, Kelly), and the practical
  constraints that decide whether any of it survives contact with a real book —
  turnover, transaction costs, cardinality, gross/net exposure, and tax-aware
  lot selection.

  Each takes its inputs positionally and by name rather than through a topic id,
  and each validates before it solves: an asymmetric covariance matrix, an
  infeasible return floor or a budget that does not sum to one is refused rather
  than quietly worked around.

Known gap

These 20 ship at contract tier — their signature is checked on every build,
but nothing here replays their arithmetic. The expected numbers exist; they sit
in each topic's tests/ directory in a shape the conformance harness does not
yet read, so this is a wiring gap and not an absence of evidence. It brings the
unasserted count to 94 of 717.

v0.13.1

Volatility and Covariance, and the diagrams that had been quietly going missing.
Nothing is removed or renamed: every subpath published in 0.13.0 resolves
unchanged, and no entry point changed its name or signature.

	0.13.0	0.13.1
Topics	675	697
Domains	17	18
Topics whose arithmetic is asserted on every build	601	623
Topics with a validated api contract	675 / 675	697 / 697
Topics carrying their diagram source	301	692

New

• Volatility and Covariance — 22 topics. Historical estimators
  (close-to-close through Yang-Zhang), realized measures (realized variance and
  covariance, bipower variation, the jump-variation detector, the realized
  kernel), conditional volatility (ARCH, GARCH, EGARCH, GJR-GARCH, FIGARCH,
  HAR-RV) and covariance estimation (sample and EWMA, Ledoit-Wolf and oracle
  approximating shrinkage, factor-model and graphical-lasso). Each ships its own
  module, and all 22 have their numbers asserted here rather than taken on trust.

Fixed

• 369 topics were publishing no diagram source. Their Mermaid is written on
  Windows and therefore CRLF, and the fence that lifts it out matched a bare
  newline — so the extraction quietly returned nothing and the page rendered
  without the diagram it was written to have. Catalog Markdown is now read with
  its line endings normalised.

Internal

• The metadata reader learned two forms of YAML it had been refusing: a block
  sequence sitting at its parent key's own indent, and a plain scalar wrapped
  across lines. Both are valid, both are what the newer generators emit, and
  between them they were dropping everything below the first wrapped summary.
• A sixth fixture convention: a named case list, where each case pins a whole
  reference output, a subset of fields, or the error code a bad input must
  raise. It is the first shape that asserts the refusals as well as the numbers.

v0.13.0

The catalog and the package finally hold the same set. 675 topics across 17
domains, with nothing skipped and nothing held back — the first release where
every algorithm that exists is also one you can install. Nothing is removed:
every subpath published in 0.12.0 still resolves.

	0.12.0	0.13.0
Topics	324	675
Domains	13	17
Topics whose arithmetic is asserted on every build	158	601
Topics with a validated api contract	324 / 324	675 / 675
Topics with an executed worked example	324	675

Technical Indicators is the headline for most people: it shipped 37 of its 137
topics in 0.12.0 and now ships all 137. ALMA, T3, VIDYA, FRAMA, TRIX, Klinger,
the Hilbert and Ehlers cycle tools and the rest are installable for the first
time.

One thing that can break you

matching-engines-and-venue-logic/continuous-matching/hybrid-pro-rata-time-matching
was publishing its family's calculate(topicId, inputs) dispatcher as its entry
point, because its slug carries a trailing word its function does not. It now
publishes hybridProRataTime(incoming, price, lot, fraction, resting).

calculate is still exported from that subpath, so an import of it keeps
working. What changed is run, which now aliases the topic's own function.
If you called run("D12-F01-A04", inputs) there, call calculate instead — or
better, the named function.

New

• Financial Mathematics, Statistics, and Data Foundations — 120 topics, from
  means and percentiles through distributions, inference, regression, and risk
  and performance statistics. This is the base layer the rest of the library is
  meant to build on, so it lives at the short subpath foundations/…. It is
  deliberately absent from the README's algorithm index, which catalogues the
  market-facing topics; it ships, it is documented, and it is in the agent skill.
• Fundamental Analysis and Valuation — 52 topics. Discounted cash flow,
  dividend discount and Gordon growth, residual income, economic value added,
  relative valuation, and the distress and earnings-quality models: Altman Z,
  Piotroski F, Beneish M, Ohlson O, Dechow-Dichev, modified Jones. Plus
  integrated and sector-specific equity scoring.
• Model Validation and Backtesting — 10 topics. ROC and PR curves, Brier
  score, log loss, calibration and expected calibration error, gains and lift,
  cost-sensitive thresholds, score migration, slice validation, rare-event bounds.
• Credit Risk and Default — 7 topics. Logistic and probit PD models,
  through-the-cycle and point-in-time PD, Merton and Bharath-Shumway distance to
  default, Campbell-Hilscher-Szilagyi distress probability.
• Geometric chart patterns gains its continuation structures and pattern
  matching families — triangles, flags, pennants, wedges, DTW and matrix profile.
• 152 topics that could not ship now do. They carried an adapter rather than
  an implementation: a file that started a Python process, piped its input across
  as JSON and returned the answer. Correct, and unusable in a package that
  promises zero dependencies and has to run in a browser. The shared runtime
  behind them is now native TypeScript, and parity was measured rather than
  assumed — 298 full-series cases and three exact EPS cases compared against the
  Python reference observation by observation, 42,435 values, worst relative
  error 1.31e-15. This is what completes Technical Indicators, and it also brings
  the Hilbert and Ehlers cycle analytics, the pivot and market-structure
  families, and the last three EPS topics.

Verification

The count nearly triples, and the description of it changes with it. These
expected values are computed in the catalog by a Python implementation written
alongside the TypeScript rather than derived from it, which makes the check
cross-language parity, not an independent third-party figure. It catches
transcription and generation errors — the failure mode that has actually
occurred here — and it would not catch both implementations sharing a misreading
of the source. The README and the agent skill now say so in those words.

Most of the jump comes from reading two fixture shapes the generator did not
recognise. 102 topics already carried a separate input and expected-output pair
on disk while nothing asserted their numbers, and the 152 ported topics carry a
richer contract again: the warm-up index, every value in the final observation,
and a causality property — that running over a prefix of the bars reproduces
exactly the prefix of the full series, which is what catches an implementation
that has quietly looked ahead. No expected value was authored to close either gap.

Contributing

src/ is still generated and still overwritten on every sync, but there are now
two routes for work that used to have none.

• optimised/ — a topic can ship a hand-written implementation from the
  repository instead of the catalog's. The catalog version is emitted beside it
  and a differential test asserts the two return identical values and throw
  identical errors, so an optimisation is reviewable by machine. npm run bench
  reports the ratio and CI fails if a recorded one regresses. The first is a
  3.5x rewrite of calculateSma.
• A proposal route for new algorithms. Assigning a topic its id stays with
  the maintainer; everything after that is a normal pull request, and it can ship
  before the article exists.

See CONTRIBUTING.md.

Fixed

• Ninety foundations topics would have published truncated subpaths —
  count-sum-minimu, look-ahead-leaka — because their folder names were capped
  at 28 characters and no slug was declared.
• Optional parameters were being dropped from the middle of published
  signatures, so a caller reading match(query, candidate, normalize) and
  passing three arguments was setting radius. Seven topics were affected.
• Arrow-form entry points recorded no parameters at all, publishing ten topics as
  zero-argument functions.
• engines.node is now >=22.12. The require condition resolves to an ES
  module and require(esm) is only unflagged from 22.12, so anyone on 22.0–22.11
  was inside the declared range and getting a hard failure.
• The bundled lookup CLI reads the docs.json beside it before reaching for the
  network, instead of answering from a different version of the library.

Internal

• A purity gate refuses any implementation that reaches a node: builtin,
  directly or through a shared module, and it now checks the whole import graph
  rather than one level. It is what caught the 152 adapter-backed topics before
  they could ship; they were ported rather than exempted, and the gate stays.
• gen-docs.mjs counts registry entries independently of the regex that parses
  them. Three times a topic had failed to match on one malformed field and
  vanished from the payload with no symptom.
• Relative imports that climb out of a topic directory are vendored into
  src/_shared/ generically, replacing a hardcoded case per engine.