chore: public-readiness pass

- Remove every reference to the internal design spec (RFC 0005 section
  numbers, draft mentions, the predict_version "spec":"0005-draft" field
  and its SQLITE_PREDICT_SPEC define): the public contract is the docs
  site; comments now describe the behavior in place
- Reword overused "honest" phrasing across README, docs site, benchmarks,
  C comments, and the PR template
- Add Dependabot config (actions, website npm, node binding, cargo, pip)
- Add bug-report and feature-request issue forms
- Add crates.io/PyPI/npm version badges to the README

Repo settings applied via API (not in this diff): de-receipted the repo
description, set homepage + topics, disabled wiki/projects, enabled
Dependabot security updates, default workflow token now read-only, and a
protect-main ruleset blocks force pushes and branch deletion.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Author: mstrathman

.github/ISSUE_TEMPLATE/bug_report.yml

@@ -0,0 +1,35 @@
+name: Bug report
+description: Something behaves incorrectly, crashes, or contradicts the docs.
+labels: [bug]
+body:
+  - type: input
+    id: version
+    attributes:
+      label: Version
+      description: Output of `SELECT predict_version()`, or the package version.
+      placeholder: v0.0.1-alpha.6
+    validations:
+      required: true
+  - type: input
+    id: platform
+    attributes:
+      label: Platform
+      description: OS/arch and how you installed (pip, npm, cargo, prebuilt binary, amalgamation).
+      placeholder: macOS arm64, pip install sqlite-predict
+    validations:
+      required: true
+  - type: textarea
+    id: repro
+    attributes:
+      label: Reproduction
+      description: The smallest SQL that shows the problem. Include table DDL and a few rows of data if the shape matters.
+      render: sql
+    validations:
+      required: true
+  - type: textarea
+    id: expected
+    attributes:
+      label: Expected vs actual
+      description: What you expected, and what happened instead (paste the full error message, they are designed to be pasted).
+    validations:
+      required: true

.github/ISSUE_TEMPLATE/feature_request.yml

@@ -0,0 +1,21 @@
+name: Feature request
+description: A capability, model, or interface you need that does not exist yet.
+labels: [enhancement]
+body:
+  - type: textarea
+    id: problem
+    attributes:
+      label: What are you trying to do?
+      description: The task in your own terms, ideally with the SQL you wish you could write.
+    validations:
+      required: true
+  - type: textarea
+    id: today
+    attributes:
+      label: What do you do today instead?
+      description: Workarounds, external tools, or "nothing, it blocks me".
+  - type: textarea
+    id: shape
+    attributes:
+      label: Proposed shape (optional)
+      description: If you have an opinion on the SQL surface or options, sketch it.

.github/dependabot.yml

@@ -0,0 +1,25 @@
+version: 2
+updates:
+  - package-ecosystem: github-actions
+    directory: /
+    schedule:
+      interval: weekly
+  - package-ecosystem: npm
+    directory: /website
+    schedule:
+      interval: weekly
+    groups:
+      site:
+        patterns: ["*"]
+  - package-ecosystem: npm
+    directory: /bindings/node
+    schedule:
+      interval: weekly
+  - package-ecosystem: cargo
+    directory: /
+    schedule:
+      interval: weekly
+  - package-ecosystem: pip
+    directory: /tests
+    schedule:
+      interval: weekly

.github/pull_request_template.md

@@ -1,4 +1,4 @@
-<!-- Thanks for contributing. Please keep this checklist honest. -->
+<!-- Thanks for contributing. Please keep this checklist accurate. -->
 
 ## What and why
 

ARCHITECTURE.md

@@ -20,7 +20,7 @@ Source layout:
 | `predict-registry.c` | the model registry (`_predict_models`) and content hashing |
 | `predict-onnx.c` | ONNX runtime backend (opt-in build only); the only file that links onnxruntime |
 | `predict-student.c` | the native student **serving** runtime: blob (de)serialization and tree / forest / MLP inference (`predict0_tree_run`). The serve side of the train/serve boundary |
-| `predict-student.h` | the shared student-model **format**: the tree/forest/MLP structs and the runtime entry points both sides agree on (RFC §4.1.3) |
+| `predict-student.h` | the shared student-model **format**: the tree/forest/MLP structs and the runtime entry points both sides agree on |
 | `predict-distill.c` | the **training** side: the CART / gradient-boosting / MLP trainers, the `distill_predict()` vtab, and the `distill_forecast()` vtab that fits the forecast student. Builds student blobs that `predict-student.c` serves |
 | `predict-internal.h` | shared types, contract constants, error codes, internal prototypes |
 | `vendor/sha256.c` | a self-contained FIPS 180-4 SHA-256 |
@@ -89,7 +89,7 @@ bytes run. Both run on CPU. A GPU build (`make loadable-onnx-gpu`,
 fp16/int8 precision; the provider-options symbols are in every onnxruntime
 C API, so it compiles and links against the CPU onnxruntime for a CI
 compile-check, while real GPU execution is validated on a dedicated GPU job.
-GPU results are honestly distinguishable from the deterministic CPU path
+GPU results stay distinguishable from the deterministic CPU path
 (provider and precision are explicit options, never silent fallbacks).
 Even so, the `benchmarks/` numbers (and the TabFM→ONNX eval in
 `benchmarks/results/tabfm-onnx.md`) are why the default answer for a
@@ -130,19 +130,19 @@ a `tree`-runtime model to the native runtime in the same file, which tells a
 single tree (`PSTREE` blob) from a forest (`PSGBT` blob) from a net (`PSMLP`
 blob) by magic and needs
 no onnxruntime. The blob formats are little-endian and normatively specified
-and versioned (RFC §4.1.3, `PSTREE01` / `PSGBT01` / `PSMLP01`), so a stored student stays
+and versioned (`PSTREE01` / `PSGBT01` / `PSMLP01`), so a stored student stays
 servable across upgrades, snapshots, and forks, and a serving-only module can
 execute a blob it never trained. They are rigorously bounds-checked on read,
 because the registry is
-writable by any SQL caller (RFC §6.2): a hand-crafted blob is rejected,
+writable by any SQL caller: a hand-crafted blob is rejected,
 never crashed on. The student is native and deterministic, like the stat
 models.
 
 Forecasting has its own student. `distill_forecast()` (`predict-distill.c`)
 fits a regression MLP that maps an instance-normalized context window of `nfeat`
 recent values to `nquant` quantiles per horizon step, distilled from a teacher's
 forecasts over sliding windows. It is stored as a fourth student blob (`PSFCST`,
-RFC §4.1.3) and served not by `predict()` but by `forecast()`
+the same registry format) and served not by `predict()` but by `forecast()`
 (`predict-forecast.c`), which extracts the most recent window, normalizes it by
 its own mean and standard deviation (so one student serves series of any
 magnitude), applies the net, and de-normalizes. When the student carries a
@@ -164,13 +164,4 @@ fp16 inference is not bit-reproducible across machines (different hardware
 and kernels round differently). Cross-machine, cross-backend bit-equality
 is never claimed (see `benchmarks/notes.md`).
 
-## Spec alignment
-
-The design spec (RFC 0005) has been amended to match the implemented
-surface: the `ts-stat`/`tabular-stat` model kinds, the aggregate calling
-convention for `forecast` and `detect_anomalies`, null option values
-meaning "key omitted," and the removal of receipts (that design is
-preserved in the RFC's Appendix D). Section references in this document
-follow the amended numbering.
-
 [tvf]: https://www.sqlite.org/vtab.html#tabfunc2

CHANGELOG.md

@@ -47,8 +47,8 @@ project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
   self-correcting error. The receipts/replay system explored during
   development (in-DB anchored receipts, then input-digest commitments,
   then document receipts) was removed before release: prediction is the
-  product, provenance was drag; the designs live in git history and the
-  RFC's deferred appendix. The model registry (`_predict_models`) remains:
+  product, provenance was drag; the designs live in git history. The
+  model registry (`_predict_models`) remains:
   content-addressed models and distilled students, hashes verified before
   deserialization.
 - **Auto model selection, conformal intervals, and `backtest()`.**
@@ -91,8 +91,8 @@ project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 - `distill_forecast()` trains a native **forecast student**: a regression net
   that maps an instance-normalized context window to future values, distilled
   from a teacher's forecasts (a foundation model run offline, or any per-window
-  teacher) over sliding windows. It registers a `PSFCST` inline BLOB (RFC 0005
-  §4.1.6) that `forecast()` serves natively, with no teacher and no onnxruntime,
+  teacher) over sliding windows. It registers a `PSFCST` inline BLOB
+  that `forecast()` serves natively, with no teacher and no onnxruntime,
   deterministically and in microseconds. The student is a **DLinear/TiDE-style
   net**: a direct linear map from the context to the fan (which carries the
   seasonal-naive + trend structure) plus a small, scaled nonlinear residual. A
@@ -131,7 +131,7 @@ project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
   rank; there is no forecast/interval. On a broad 200-series TSB-AD-U sample it
   scores 0.44 median VUS-PR (at the published SOTA level) versus the theta
   residual detector's 0.23 -- roughly 2x on the typical series, all in zero-
-  dependency C. Two honest caveats: the *mean* gap is
+  dependency C. Two caveats: the *mean* gap is
   much smaller (0.41 vs 0.38) because sub-pca fails badly on weakly-periodic or
   very long series where the fixed-window PCA does not fit, so the two are partly
   complementary; and on a subset filtered to moderate-length periodic series the

README.md

@@ -2,6 +2,9 @@
 
 [![CI](https://github.com/PureStorage-OpenConnect/sqlite-predict/actions/workflows/ci.yml/badge.svg)](https://github.com/PureStorage-OpenConnect/sqlite-predict/actions/workflows/ci.yml)
 [![License: MIT OR Apache-2.0](https://img.shields.io/badge/license-MIT%20OR%20Apache--2.0-blue.svg)](#license)
+[![crates.io](https://img.shields.io/crates/v/sqlite-predict.svg)](https://crates.io/crates/sqlite-predict)
+[![PyPI](https://img.shields.io/pypi/v/sqlite-predict.svg)](https://pypi.org/project/sqlite-predict/)
+[![npm](https://img.shields.io/npm/v/sqlite-predict.svg)](https://www.npmjs.com/package/sqlite-predict)
 ![C99](https://img.shields.io/badge/C99-zero%20dependencies-brightgreen.svg)
 
 **Prediction as a SQL primitive.** `forecast()`, `detect_anomalies()`, and
@@ -32,7 +35,7 @@ that lives inside your database and runs anywhere.
 
 ## Is it accurate?
 
-Yes, and the numbers are honest and reproducible (`benchmarks/`), not a demo on
+Yes, and the numbers are measured and reproducible (`benchmarks/`), not a demo on
 a toy series.
 
 - **Forecasting.** A Chronos foundation model distilled into a zero-dependency
@@ -166,7 +169,7 @@ SELECT avg(coverage) FROM backtest('SELECT ts, value FROM readings', 6,
 ## Models
 
 The default build is **pure C with no dependencies**. It ships small,
-honest statistical models:
+well-studied statistical models:
 
 - `theta-classic` (the Theta method) and `stub-seasonal-naive` for `forecast()`
   and `detect_anomalies()`; `tsb` (Teunter-Syntetos-Babai) for

benchmarks/compare_tabular.py

@@ -2,7 +2,7 @@
 
 Candidates: Google TabFM (local weights under models/tabfm, downloaded
 from google/tabfm-1.0.0-pytorch — license tabfm-non-commercial-v1.0,
-benchmark/evaluation use), plus honest floors (majority class / global
+benchmark/evaluation use), plus baseline floors (majority class / global
 mean) and a pure-python 5-NN so wins are attributable. The extension's
 predict() column joins when M4 lands. Unavailable candidates are
 reported SKIPPED, never silently dropped.

benchmarks/forecast_distill.py

@@ -24,9 +24,9 @@
 GENERICITY. Every hyperparameter is a fixed function of the seasonal period M
 (R=2M recent lags, min-origin 3M, stride M/2), and M comes from dataset
 metadata, not from tuning. The SAME pipeline runs on all three regimes. This is
-the honest test: on m4_hourly the FM beats the stat floor, so distillation
+the real test: on m4_hourly the FM beats the stat floor, so distillation
 should help; on tourism_monthly the FM does not beat our stat model (see
-forecast.md), so an honest method must NOT manufacture a win there. Nothing is
+forecast.md), so a sound method must NOT manufacture a win there. Nothing is
 selected on the test window.
 
 FM forecasts are cached per (dataset, series, origin) under

benchmarks/forecast_native.py

@@ -4,7 +4,7 @@
 This is the end-to-end proof that a *capable* student recovers most of a
 forecast FM's edge, where a tree student could not. Chronos labels sliding
 context windows (offline, cached); `distill_forecast` fits a multi-output
-regression MLP (PSFCST, RFC 0005 §4.1.6) that reproduces the teacher; then
+regression MLP (a PSFCST blob) that reproduces the teacher; then
 `forecast()` serves that native student with no teacher and no onnxruntime.
 
 On m4_hourly (80 series, context 256, horizon 48):