Precompilation v2: compiled form, Heddle.Build, generated sites (3.0.0)

.claude/rules/api-compatibility.md

@@ -0,0 +1,18 @@
+---
+paths:
+  - "src/Heddle/**/*.cs"
+  - "src/Heddle.Tool/**/*.cs"
+---
+
+# API design and compatibility
+
+Details: [coding-standards.md § API](../../docs/spec/common/coding-standards.md#api-design-and-compatibility), [breaking-windows.md](../../docs/spec/common/breaking-windows.md). Anchored in the .NET Framework Design Guidelines.
+
+- Additive by default: new extensions, opt-in options, overloads. Existing templates and hosts compile and render byte-identically.
+- Breaking changes land only in a ratified breaking window (D2), each with a migration note. Candidates go to the register in breaking-windows.md, not into code.
+- Every new `TemplateOptions` property goes into the copy constructor (and `Equals`/`GetHashCode` when it keys identity) — missed before (`ProvideLanguageFeatures`); the completeness test enforces this.
+- Public API additions carry XML docs + same-change updates to the affected published docs pages (`language-reference.md`, `custom-extensions.md`, `built-in-extensions.md`, `csharp-api.md`).
+- Extend via the sanctioned seams (`[ExtensionName]` registry, options, function registry, output profiles, publish/read channel, resolver, encoder) — don't invent parallel mechanisms or new seams for single-implementation internals.
+- The build names no extension: no compare against a built-in name in any form, no `"Heddle.Extensions."` type literal, no `"Heddle"` assembly-name compare, no name-keyed table. Read the binder, the attributes, or the hook.
+- The build never predicts a compile-time hook either: the loader runs the extension's real `InitStart`/`CompleteInit` at materialization (once per request shape, over bodies served from the compiled form), so a bodied, hook-overriding, `[Prop]`-declaring or `[BranchRole]` custom extension precompiles in a default build. A call the build cannot serve costs **that template** (an `HED7031` decline with a machine-readable class), not the build — reach for a build error only where the template must not precompile silently.
+- Extension instances are shared across concurrent renders: no mutable per-render state on them; per-render state lives in the `Scope` lineage.

.claude/rules/coding-style.md

@@ -0,0 +1,18 @@
+---
+paths:
+  - "**/*.cs"
+---
+
+# Coding style
+
+Authority order: (1) surrounding code in the file, (2) [coding-standards.md](../../docs/spec/common/coding-standards.md), (3) `.editorconfig` (editor guidance only, not a gate).
+
+- Block-scoped namespaces (`namespace X { … }`) — not file-scoped, even in new files.
+- Allman braces predominantly — match the file. Braceless single-statement guards for early returns are idiomatic; multi-line/nested bodies always braced.
+- `var` when the type is apparent; explicit types otherwise and for built-ins.
+- Naming: `_camelCase` private instance fields; `PascalCase` private static readonly; `I`-prefixed interfaces; extension classes `<Name>Extension` with `[ExtensionName("name")]`.
+- Existing quirks (space before parameter list, space after cast) are deliberate — do not "fix" them.
+- Comments follow C1–C8: default to none; write one only when the code cannot carry the meaning; never cite specs, plans, findings, or reviews in code comments (C4). XML doc comments on public API; regression tests carry a doc comment explaining the pinned scenario.
+- 4-space indent, UTF-8, final newline, ~120-column wrap.
+- Never mass-reformat or run whole-file IDE cleanups; diffs contain only the lines the change needs. The established style is the maintainer's deliberate choice — consistency beats modernization.
+- `src/Heddle.Language/generated/` is generated (ANTLR) — never hand-edited; grammar change = `.g4` edit → regen → commit both.

.claude/rules/errors-diagnostics.md

@@ -0,0 +1,17 @@
+---
+paths:
+  - "src/Heddle/**/*.cs"
+  - "src/Heddle.Tool/**/*.cs"
+  - "src/Heddle.Language*/**/*.cs"
+---
+
+# Errors and diagnostics
+
+Details: [coding-standards.md § errors](../../docs/spec/common/coding-standards.md#error-handling-and-diagnostics), [cross-cutting-decisions.md D1 + registry](../../docs/spec/common/cross-cutting-decisions.md).
+
+- Compile-path problems are collected, not thrown: positioned `HeddleCompileError`/`HeddleCompileWarning` on the compile result, so template authors see all problems at once. The public API never surfaces raw exceptions for template mistakes.
+- Host-programming errors throw at public API entry points (`ArgumentNullException` etc.).
+- Every new diagnostic claims a stable `HED####` ID from the registry (feature-area blocks, D1) **in the same change**; IDs are never reused or renumbered.
+- An ID whose condition becomes unreachable is **retired in place**, not deleted: the `HeddleDiagnosticIds` constant, the `HeddleDiagnosticCatalog` row, the registry row and a mention in the page the registry names as owner all stay, and the registry row says what stopped firing, why, and what replaced it (often nothing). `DiagnosticIdTests` gates all four; deleting any would free the number for a later, different fault.
+- Messages name the construct, point at the remedy, and carry the template position.
+- Security-sensitive failures never degrade to execution: a sandbox violation is a compile error, full stop.

.claude/rules/performance.md

@@ -0,0 +1,16 @@
+---
+paths:
+  - "src/Heddle/**/*.cs"
+  - "benchmarks/dotnet/**"
+---
+
+# Performance (render path is hot)
+
+Details: [coding-standards.md § performance](../../docs/spec/common/coding-standards.md#performance-rules-the-render-path-is-hot).
+
+- No new per-render allocations. `Scope` stays a `readonly struct` passed by `in`; optional per-render state is lazily created.
+- Reflection at compile time only; render time executes pre-compiled delegates and direct calls.
+- `[MethodImpl(AggressiveInlining)]` only on tiny, provably-hot transforms — with a benchmark justifying it.
+- Perf claims are proven by BenchmarkDotNet (`benchmarks/dotnet`), never asserted. Hot-path changes run affected benchmarks before/after on the same machine: allocated bytes must not increase; mean must stay within reported error. Intentional trade-offs need maintainer ratification.
+- Compile-time cost is "compile once, render many" — moderate compile-path allocation is fine, but startup/precompilation paths get measured too.
+- No JIT-rescue assumptions for escaping allocations (D7).

.claude/rules/spec-workflow.md

@@ -0,0 +1,15 @@
+---
+paths:
+  - "docs/spec/**"
+---
+
+# Spec and decision workflow
+
+Details: [spec-conventions.md](../../docs/spec/common/spec-conventions.md), [cross-cutting-decisions.md](../../docs/spec/common/cross-cutting-decisions.md) (which also carries the condensed release and program records).
+
+- Specs live under `docs/spec/`, indexed from `docs/spec/README.md` (update the index in the same change); they are contributor material, never published (D9).
+- Plan = what/why (ratified input); spec = exactly how. A finished spec has **no open questions** — every point is a closed decision with evidence, or a most-reversible default plus a named revisit trigger.
+- Implementation follows the owning plan's declared item order (D5): assume 1..N−1 merged; never depend on later items.
+- Changing an earlier recorded decision goes through the amendment mechanism (evidence → maintainer ratification → implemented by the amending effort); earlier spec text is never silently retrofitted. The accumulated E1–E28 ledger is closed — new amendments are recorded as dated notes in the owning spec (see cross-cutting-decisions.md § Cross-spec amendments ledger).
+- Balanced-mode principle precedence on conflicts: correctness/back-compat → sandbox security → simplicity → measured hot-path performance → abstraction/extensibility. DRY consolidates knowledge (rule of three); YAGNI cuts presumptive features, never tests/benchmarks/docs.
+- Writing style: sentence-case headings, en/em dashes, relative links only, cite `path` + member name over bare line numbers.

.claude/rules/testing.md

@@ -0,0 +1,22 @@
+---
+paths:
+  - "src/*Tests*/**"
+  - "src/TestCorpus/**"
+  - "samples/**"
+---
+
+# Testing
+
+Details: [testing-standards.md](../../docs/spec/common/testing-standards.md) — normative; this is the short form.
+
+- Loop: write the spec's tests failing first → implement to green → run the gate → fix forward. Goldens/baselines/spec tables change only with maintainer review, never to silence a red test.
+- Gate (one combined run): `dotnet build -c Release` → `dotnet test --project src/Heddle.Tests/Heddle.Tests.csproj` (all TFMs, serially; both Debug and Release when `#if DEBUG` arms are touched) → no diff under `src/Heddle.Language/generated/` (unless the spec licenses a grammar change) → benchmarks when a hot path is touched → `npm run docs:build` when docs pages changed.
+- xUnit v3 on MTP: `dotnet test --project <csproj>` is the canonical form; filters are MTP syntax after `--` (`--filter-method` etc.); a filter matching nothing exits 8; CI legs run through `.github/scripts/dotnet-test-guarded.sh <project>`, which passes `--fail-skips on`.
+- **Run a suite once and keep the whole output.** These suites take minutes; re-running one to see a detail you piped away is the single most expensive habit available. Always `2>&1 | tee` (or redirect) the full run to a file, then read that file — grep it for failures, open the failing case, count tests, whatever is needed. Never pipe a run through `tail`/`grep` as the only sink, and never re-run a suite to answer a question the first run already contained. Redirect per suite so runs stay separable, and prefer one combined script that captures every suite in a single pass over four separate invocations.
+- **No test-count floors.** `--minimum-expected-tests` was tried on nine CI legs and withdrawn: it is satisfied by editing a digit, it names nothing when it reddens, it drifted (one commit raised `dotnet.yml` and left `lsp.yml`, and that leg then tolerated a six-test regression for four commits), and it fails a leg for the wrong reason whenever a test is legitimately quarantined. What a suite CONTAINS is gated inside the suite: `src/<Suite>/test-classes.txt`, one line per fact-declaring class, set-equality asserted by that suite's `TestClassInventoryTests` — a red gate names the class. Deliberately the same shape as corpus membership below.
+- Fixtures: `.heddle` + sibling golden under `src/Heddle.Tests/TestTemplate`; every new golden/fixture directory is LF-pinned in `.gitattributes` **in the same change**.
+- Negative/security tests assert positioned `HED*` diagnostics (ID + position) and prove the construct never executes.
+- Descriptive PascalCase sentence test names; assert with context (`Assert.True(result.Success, result.ToString())`); tests are exempt from DRY.
+- Concurrency: any state shareable across renders ships a parallel-render isolation test.
+- Precompiled tier: fallback to dynamic is byte-identical, so an unpinned test proves nothing about the tier it claims to test. End-to-end tests pin the tier (`PrecompiledMismatchPolicy.Strict` + fallback sentinel); expected fallbacks are declared only via `FallbackGuard.Expect(key, reason)` (run-time) / `CompiledFormHarness.ExpectRefusal(key, class)` (build-time decline beyond the corpus row); quarantine guarded fixtures with an owning-item `Explicit = true` (reported *not run*, so it survives `--fail-skips`), never weaken them.
+- Test-input single-sourcing: a template shape verified by more than one tier exists once, in the shared corpus, with declared (total) intent — duplicate copies drift silently and break the differential premise. Membership is gated by set equality — never a count or floor (a floor once hid 15 templates silently dropping out).

.gitattributes

@@ -13,15 +13,41 @@
 # output — failing on Windows while Linux (LF everywhere) passes. Pin every such root to LF so the
 # byte comparison holds identically on both platforms regardless of the developer's autocrlf setting.
 # All files under these roots are text (verified: no binaries). New golden/fixture roots MUST be
-# added here in the same change that introduces them — see docs/spec/common/testing-standards.md.
+# added here in the same change that introduces them.
+#
+# eol=lf pins line endings and says nothing about byte-order marks; the two are independent, and a
+# BOM drift can pass every eol=lf check there is. Templates carrying a deliberate UTF-8 BOM declare
+# it in src/TestCorpus/CorpusIntent.cs, gated by CorpusIntentGateTests.
 samples/**/golden/**                              text eol=lf
 src/Heddle.Tests/TestTemplate/**                  text eol=lf
+src/Heddle.Tests/TestOutput/**                    text eol=lf
 src/Heddle.Tests/Streaming/**                     text eol=lf
-src/Heddle.Generator.Tests/Snapshots/**           text eol=lf
-src/Heddle.Generator.IntegrationTests/Fixtures/** text eol=lf
+src/Heddle.Build.Tests/Fixtures/**             text eol=lf
 src/Heddle.LanguageServices.Tests/Corpus/**       text eol=lf
 src/Heddle.Demo.Wasm/contract-fixtures/**         text eol=lf
 
+# The per-suite test-class inventories are read line by line and compared as a set on both platforms.
+# Reading normalises CRLF, but pinning them keeps the diff a reviewer sees identical everywhere.
+src/*/test-classes.txt                            text eol=lf
+
+# Golden oracle corpus: the *.golden.html files intentionally end WITHOUT a trailing newline and must
+# round-trip byte-exact on every platform, so they are pinned -text (no smudging at all) rather than
+# eol=lf. The JSON sidecars are ordinary LF text.
+benchmarks/dotnet/GoldenCorpus/*.golden.html -text
+benchmarks/dotnet/GoldenCorpus/*.verify.json text eol=lf
+benchmarks/dotnet/GoldenCorpus/manifest.json text eol=lf
+# Exported model fixtures are hash-recorded in the manifest like the goldens; LF text.
+benchmarks/dotnet/GoldenCorpus/fixtures/**   text eol=lf
+
+# Rust benchmark harness: composed-page fragment data must round-trip byte-exact (-text); template
+# files are ordinary LF text (the harness erases their line endings, but LF keeps diffs clean).
+benchmarks/rust/templates/** text eol=lf
+
+# Python benchmark harness: template files are committed LF on every platform — Mako passes a literal
+# \r\n in a template file through into rendered output, so LF-pinning keeps checked-in bytes equal to
+# authored bytes.
+benchmarks/python/templates/** text eol=lf
+
 # Custom for Visual Studio
 *.cs     diff=csharp
 *.sln    merge=union
@@ -41,3 +67,6 @@ src/Heddle.Demo.Wasm/contract-fixtures/**         text eol=lf
 *.PDF	 diff=astextplain
 *.rtf	 diff=astextplain
 *.RTF	 diff=astextplain
+
+# The stored compiled-form fixture is bytes, not text: never normalised (P4-R3).
+src/Heddle.Tests/TestTemplate/*.bin  binary

.github/rulesets/main-branch-protection.json

@@ -31,6 +31,7 @@
         "required_status_checks": [
           { "context": "build (ubuntu-latest)" },
           { "context": "build (windows-latest)" },
+          { "context": "release-tests / suites" },
           { "context": "dco" }
         ]
       }

.github/scripts/dotnet-test-guarded.sh

@@ -0,0 +1,38 @@
+#!/usr/bin/env bash
+# Runs one test project under Microsoft.Testing.Platform, with the two guards that are worth having.
+#
+# Usage: dotnet-test-guarded.sh <project> [extra dotnet test args...]
+#
+# `--fail-skips on` turns a skipped test into a failed one. A skip is a test that stopped testing, and
+# CI is the wrong place to find out quietly; the repository's protocol for a known-failing test is
+# `[Fact(Explicit = true)]`, which is reported as NOT RUN rather than skipped and so survives this.
+#
+# Exit 8 -- a filter matched nothing, or discovery failed -- stays called out by name, because a stale
+# filter that runs nothing used to pass. It is the platform's own signal, not a number anyone maintains.
+#
+# What this script no longer does is carry a per-suite `--minimum-expected-tests` floor. Nine of them
+# were maintained by hand across two workflows and they failed at their own job: one commit added six
+# tests, raised the floor in one workflow and left the other, and that leg then tolerated a six-test
+# regression in silence for four commits. A floor is satisfied by editing a digit, it names nothing
+# when it reddens, and it fails a leg for the wrong reason whenever a test is legitimately quarantined.
+# Suite membership is gated instead by the checked-in test-class inventories (src/<Suite>/test-classes.txt,
+# asserted by set equality inside each suite), which name the class that appeared or vanished.
+
+set -uo pipefail
+
+if [ "$#" -lt 1 ]; then
+  echo "usage: $0 <project> [extra args...]" >&2
+  exit 64
+fi
+
+project=$1
+shift 1
+
+dotnet test --project "$project" "$@" -- --fail-skips on
+status=$?
+
+if [ "$status" -eq 8 ]; then
+  echo "::error::Zero tests ran in ${project} -- a filter matched nothing, or discovery failed."
+fi
+
+exit "$status"