simt-dev 0.2.0__py3-none-any.whl

simt_dev/vendored/docs/decisions/ADR-022-a-key-simt-dev-reads-is-prefixed.md

@@ -0,0 +1,98 @@
+# ADR-022 — A key simt-dev reads is prefixed `simt-`
+
+Status: accepted · 2026-10-08
+**Affects:** `rules/*.md`, `rules/mdc/*.mdc`, `src/projection/mdc.mjs`, `src/projection/rules.mjs`, `test/shipped.test.mjs`. Supersedes the *spelling* of the keys named in ADR-013 and ADR-016; their rulings stand.
+
+## Context
+
+A canonical `.mdc` frontmatter block carries two vocabularies in one place:
+
+```
+id: test-gate
+target: .cursor/rules/test-gate.mdc
+description: ...
+alwaysApply: true
+applies: always
+```
+
+`description` and `alwaysApply` are Cursor's, cited in the `cursor` row of
+`src/projection/surfaces.mjs` and re-emitted per surface. `id`, `target`, `applies`,
+`optIn` and `substitutions` are simt-dev's own and appear in no vendor's
+documentation. In the file above, nothing says which is which: `alwaysApply: true` and
+`applies: always` sit four lines apart, both read as "always", and both are consumed by
+something. The question was asked of this repository and could only be answered by
+reading the source rather than the file.
+
+That is a legibility failure with teeth behind it. ADR-016 records finding S11: a rule
+shipped, was tested for its presence on disk, and reached no repository, because
+`applies` was absent from its frontmatter. A missing simt-owned key is silent. A bare
+key that looks like the key beside it is one typo away from the same silence, and this
+block is precisely where a reader editing one key is looking at the other.
+
+The prefix does not exist to make the files tidy. It exists so that a bare key is
+either an editor's key or a defect, and never a third thing.
+
+## Rulings
+
+1. **Every key simt-dev reads out of a canonical rule file carries `simt-`.**
+   `simt-id`, `simt-target`, `simt-title`, `simt-applies`, `simt-opt-in`,
+   `simt-substitutions`, `simt-fallback`.
+
+   | was | is |
+   |---|---|
+   | `id`, `target`, `title` (`.mdc`) | `simt-id`, `simt-target`, `simt-title` |
+   | `applies`, `optIn`, `substitutions` (`.mdc`) | `simt-applies`, `simt-opt-in`, `simt-substitutions` |
+   | `id`, `title`, `applies`, `fallback` (`rules/*.md`, `.simt-dev/profiles/`) | `simt-id`, `simt-title`, `simt-applies`, `simt-fallback` |
+
+2. **An editor's key keeps the vendor's spelling, unprefixed.** `description` and
+   `alwaysApply` are the only two keys in the estate with two owners: simt-dev reads
+   them and re-emits them mapped per surface (`alwaysApply` becomes Antigravity's
+   `trigger: always_on`). Prefixing them would be a lie in both directions — Cursor
+   would ignore `simt-alwaysApply`, and the projector would have to strip the prefix
+   before emitting. A key is prefixed when this tool is its only reader.
+
+3. **An unrecognised key is a defect, not an ignored line.** `applies:` in a file that
+   should say `simt-applies:` reports as a problem naming the prefixed spelling.
+   This is the ruling that turns the prefix from a convention into an invariant: the
+   failure mode ADR-016 hit is now loud, and loud before the rule reaches nobody.
+
+4. **The prefix binds settings read from files, not identifiers in code.** `rule.applies`
+   is still `rule.applies` in `src/`. Renaming the property would satisfy the sentence
+   and change nothing a reader of a rule file sees, which is the only reader the prefix
+   is for.
+
+5. **Inside `simt-substitutions:` the keys stay `key`, `default` and `describe`.** The
+   parent already namespaces them; `simt-key` reads as though the block were not there.
+
+6. **Out of scope, deliberately: `simt-dev.json` and `graph/skills.graph.json`.** Both
+   are namespaced by filename, neither contains a key any editor or agent reads, and
+   both are read by this tool's own commands rather than by a person opening a rule
+   file. The receiving repo's manifest is the one file a human edits by hand — to
+   declare a gate command or a substitution — and `simt-gates` there is noise in the
+   place noise costs the most.
+
+## Consequences
+
+- **No receiving repository is affected.** These keys are stripped before projection:
+  `.md` frontmatter is dropped by `stripFrontmatter`, and a projected `.mdc` gets
+  frontmatter rebuilt from the surface row (`description` + `alwaysApply`) or preserves
+  whatever an adopted file already had, verbatim and unparsed. The change is
+  confined to the files in this repository, which is why touching shipped rule text is
+  cheap here and expensive in `rules/core.md`'s prose.
+- A rule file that still says `applies:` now fails `check` with the prefixed spelling in
+  the message. That is the intended break: it fails at the file, loudly, on first run.
+- ADR-013 and ADR-016 name the old spellings in their examples. They are left verbatim
+  — they are the record of what was ruled — and this record is the lookup from the old
+  spelling to the new one.
+- `rules list` keeps an `applies` column. It labels a value, not a key, and the value
+  vocabulary (`always`, a tag list) is unchanged.
+
+## Rejected
+
+- **Prefix the manifest keys too.** Rejected: `simt-dev.json` is already namespaced by
+  filename, and the keys a person edits by hand are the ones a prefix would obscure.
+- **Prefix in code as well, for symmetry.** Rejected under ruling 4: it is a rename of
+  variables that satisfies the wording and defeats the purpose.
+- **Accept both spellings for one release.** Rejected: a key that may be written two
+  ways is exactly the ambiguity this record exists to remove, and the failure it hides
+  is the unreachable rule ADR-016 already shipped once.

simt_dev/vendored/docs/decisions/ADR-023-the-gate-ships-as-a-file-the-repo-owns.md

@@ -0,0 +1,106 @@
+# ADR-023 — The gate ships as a file the repo owns
+
+Status: accepted · 2026-10-06 · ruling 2 superseded by
+[ADR-024](ADR-024-simt-dev-overwrites-its-own-gate-file.md), which carves out this
+repository's own copy; rulings 3 and 4 narrowed by the same record to every repository
+except this one. The text below is left as it was accepted.
+**Affects:** `src/ci/workflow.mjs`, `src/locate.mjs`, `src/hooks/shim.mjs`, `src/commands/init.mjs`, `src/projection/project.mjs`, `test/ci.test.mjs`, `test/leaks.mjs`. Settles plan 1 §6 phase 3 ("Reusable workflow live").
+
+## Context
+
+`simt-dev check` is the whole gate: projection drift, an unreadable rule, an unrunnable
+gate, a stale manifest. `init` installs it as a `pre-commit` hook, which means it runs on
+the machine of whoever remembers to run it, and a hook is a developer's machine. Every
+review is somebody else's, and every merge is a machine nobody gated.
+
+The plan called this "a reusable workflow", and named
+`the organisation/simt-dev/.github/workflows/sync-check.yml` as a file "referenced by every
+consumer". Nothing in this repository had a `.github/` at all. There were three ways to
+close the gap, and they differ in who the resulting file belongs to:
+
+- **`uses:` a reusable workflow.** One file, changed centrally, no copying. Every consumer
+  references a path inside a repository its team has no access to, cannot diff in review,
+  and cannot edit. The gate's own behaviour becomes a function of someone else's push, and
+  a red build becomes something the reader cannot diagnose from what they have.
+- **Copy the bytes in.** Each repo gets a file it can read, edit, argue with and delete.
+  Improving the template does not reach repos that already have it.
+- **Run the tool from the consumer's CI, configured by them.** No shipped file at all,
+  which means every repo writes the same three lines and none of them are checked.
+
+This repository has one rule that settles it: **every byte that leaves this repository
+must be readable and actionable by someone who has only the receiving repository.** A
+`uses:` line points at bytes the reader does not have. That is not a style preference; it
+is the failure this tool exists to stop having shipped, and it has shipped — the projected
+rules once told readers to edit a source tree they did not have.
+
+## Rulings
+
+1. **The gate ships as copied bytes, written into `.github/workflows/simt-dev-check.yml`.**
+   A `uses:` reference is ruled out by the rule above. The file is fully readable in the
+   pull request that adopts it, which is the only review it ever gets.
+
+2. **It is written once and never rewritten.** From creation the file belongs to the repo:
+   their extra jobs, their triggers, their matrix. Overwriting it would delete work this
+   tool has no way to know about, which is the same defect as replacing a hand-written
+   `.mdc` body (ADR-016). `planWorkflow` reports an existing file as `unchanged` and never
+   reads its contents for comparison.
+
+   **The cost, stated rather than discovered later:** a template improvement does not
+   reach an adopted repo. There is no marker, no managed region, no way for `upgrade` to
+   tell their edit from ours. A repo that wants the newer template takes the diff by hand.
+   This is a deliberate trade, and ruling 4 is what keeps it cheap.
+
+3. **The gate only appears where GitHub Actions already runs.** `.github/workflows/` must
+   exist as a directory. Creating it silently would start running jobs on pull requests the
+   team did not ask for, which is the mirror image of the defect: a tool making a
+   decision for someone and reporting success. `init` prints `skip` with the reason and
+   the way out — "create `.github/workflows/` first if this repo is on GitHub Actions,
+   then rerun init" — so a skip is a decision they can reverse rather than an absence they
+   discover from a failing build.
+
+4. **`check` reports the absence; it never inspects a file's contents.** An Actions repo
+   with no simt-dev workflow is a blocking `absent` finding with the remedy
+   `simt-dev init --yes`. A workflow that exists is never diffed, never opened, never
+   counted. A gate that inspects a file it also owns would report its own drift forever,
+   and the reader's only fix would be to re-run `init` — the "permanently red check
+   nothing can fix" failure the code comments fear elsewhere. Their file, their content.
+
+5. **The binary resolution is one function, shared with the hook.** The hook shim already
+   resolved the dispatcher through `$SIMT_DEV_BIN` → `node_modules` → `PATH` → refuse. That
+   block moved to `src/locate.mjs` so the workflow could take it rather than re-spell it.
+   Two copies diverge the first time one is fixed, and the symptom is the worst available:
+   a gate that passes on a laptop and cannot start in CI.
+
+6. **A job that cannot find the tool fails.** Every resolution chain ends in a refusal with
+   a non-zero exit, and the file says so in its own header. A green tick from a check that
+   never ran is the one outcome worse than a red one.
+
+## Consequences
+
+- `init` writes one more file, so one more line in its plan. Every file `init` creates is
+  named in the plan it prints: a file the reader was not shown is a file they approve by
+  reading the diff afterwards, which is the one review step this tool exists to make
+  unnecessary.
+- The rendered bytes are subject to the same leak check as rule bodies, now shared as
+  `test/leaks.mjs`. This is the first thing simt-dev ships that a human has to *read and
+  argue with* in a review rather than merely receive, so the guard is a requirement and not
+  a courtesy.
+- The workflow's `run:` body is checked with `sh -n`. A shell syntax error inside a YAML
+  block scalar is otherwise found on the adopting repo's first pull request, in public.
+- A repo with no `.github/` sees no change at all: no finding, no file, no directory.
+- The template is not a managed region and this record does not pretend otherwise. If a
+  future release changes what the gate does, existing repos keep what they have.
+
+## Rejected
+
+- **`uses: the organisation/simt-dev/.github/workflows/sync-check.yml@main`.** Rejected under
+  ruling 1: it is the plan's own suggestion, and it points every reader at bytes they do
+  not have. Worth recording that the plan wrote it, because the next person to read §6
+  will find it there.
+- **Maintain it as a marker-managed region, refreshed by `upgrade`.** Rejected: it needs a
+  merge strategy for a repo that added their own job, which is the problem ruling 2
+  declines to take on. A refreshed file that overwrites a deployment job is worse than a
+  stale one.
+- **Create `.github/workflows/` when it is missing.** Rejected under ruling 3.
+- **Ship a reusable workflow *and* copy it.** Rejected: two sources for one gate, and the
+  copy is the one nobody reads.

simt_dev/vendored/docs/decisions/ADR-024-simt-dev-overwrites-its-own-gate-file.md

@@ -0,0 +1,111 @@
+# ADR-024 — simt-dev overwrites its own gate file, and no repository else's
+
+Status: accepted · 2026-10-06
+**Affects:** `src/ci/workflow.mjs`, `src/commands/ci.mjs`, `src/commands/init.mjs`,
+`src/commands/index.mjs`, `src/commands/help.mjs`, `src/projection/project.mjs`,
+`.github/workflows/simt-dev-check.yml`, `test/ci.test.mjs`. Supersedes ruling 2 of
+[ADR-023](ADR-023-the-gate-ships-as-a-file-the-repo-owns.md), and narrows its rulings 3 and
+4 to every repository except this one. ADR-023's rulings 1, 5 and 6 stand.
+
+## Context
+
+ADR-023 copied the gate into `.github/workflows/` and wrote it once. For an adopting
+repository that is the right call, and the cost was written down rather than discovered
+later: a template improvement does not reach a repo that already has the file.
+
+This repository is not an adopting repository. It is the tool. And a tool whose own gate is
+a frozen copy of its own gate is a tool with two gates, one of which it does not maintain.
+The file `.github/workflows/simt-dev-check.yml` in this repository was written by the
+template, and every future change to the template would leave it behind — silently, because
+ADR-023 ruling 4 says `check` never inspects a workflow's contents, so nothing would report
+the gap. A gate that cannot report its own staleness is the failure mode ADR-017, ADR-018
+and ADR-019 were each written about.
+
+There is a second reason the exception is narrow rather than general. **A rule with one
+exception is a rule with an exception; a rule with a flag is a rule with two behaviours
+depending on who asked.** The write-once ruling is what protects a repo's deployment job
+from being overwritten. "Unless the tool feels like it" is not a boundary anyone can check.
+The only boundary that needs no judgement is a path, and this repository is one path.
+
+## Rulings
+
+1. **This repository's gate file is generated, refreshed, and generated bytes only.** It
+   lives at `PACKAGE_ROOT`, recognised by `resolve(root) === PACKAGE_ROOT` — not by a
+   flag, an env var, or a name in a config file, because all three can be set by a consumer
+   who would then have found the one path that overwrites their CI. `resolve`, not
+   `realpath`: a symlinked checkout that fails to match leaves it a consumer, which is the
+   direction a mistake is survivable in.
+
+2. **`simt-dev ci` is the only writer of that copy, and it is a dry run without `--yes`.**
+   `init` cannot be the writer, and this is not a design preference. `init` on this
+   repository would pin this repository's SHA into its own `AGENTS.md`, which is precisely
+   what the "Why this repository is not self-adopted" section of `AGENTS.md` exists to
+   prevent. A refresh reachable only through `init` is an exception that can never fire, and
+   an exception that cannot fire is worse than no exception: it reads in the code as though
+   the copy were maintained.
+
+3. **`simt-dev ci` outside this repository is a refusal, not an error.** It prints who owns
+   the file and the command that owns it (`simt-dev init --yes`), writes nothing, and exits
+   zero. A refusal is a report. Making it a failure would mean a CI job that runs `ci`
+   before it runs `check` breaks for every adopting repo, for a condition that is correct.
+
+4. **The header of each copy says which world it is in.** The consumer copy says it is
+   written once and never rewritten. The own copy says it is refreshed, that a diff means
+   the **gate** changed rather than that somebody hand-edited their CI, and that the way to
+   change it is `src/ci/workflow.mjs` in the same commit followed by `simt-dev ci --yes`.
+   Both sentences are tested against the rendered bytes. The alternative is a reader of this
+   repository's diff seeing their CI change and having to guess who did it.
+
+5. **This repository's copy is compared; every other repository's is not.** `planWorkflow`
+   computes `current` only when the root is this repository's, and `check` raises a
+   `drift` finding with the remedy `simt-dev ci --yes` only then. ADR-023 ruling 4 stands
+   unchanged for everyone else: their file, their content, never opened.
+
+6. **This repository's copy runs the gate this repository can actually pass.** The consumer
+   copy runs `simt-dev check`, which fails here by design — `check` reports
+   `UNMANAGED AGENTS.md` in a repository that has deliberately not adopted itself, so a copy
+   of the consumer workflow committed here would be red on its first run and every run after.
+   The own copy runs what `.githooks/pre-push` already runs — `npm test` and
+   `graph lint` — plus `simt-dev ci` itself, so the file is its own staleness check. A test
+   asserts both commands are in the hook and in `package.json`, because a gate that names a
+   command this repository does not have is ADR-019 again.
+
+7. **Both copies render from one skeleton.** Triggers, permissions, concurrency, runner and
+   timeout are written once in `skeleton(name, steps)`; the two renderings differ in the
+   header text, the job name and the steps. Two independently written YAML files are two
+   opinions about when a pull request is gated, and the copy that is wrong is the copy nobody
+   diffs.
+
+## Consequences
+
+- `simt-dev ci` is a new command with one job. It is also the first command in this tool
+  that only makes sense in the tool's own repository, and that is deliberate: it is the
+  only honest way to ship a self-maintained generated file without giving the write-once
+  ruling a flag.
+- `test/ci.test.mjs` now has one test that writes into this repository: it mutates the
+  committed workflow, asserts `check` reports the drift, runs `ci --yes`, asserts the bytes
+  are restored, and restores in a `finally`. No fixture can produce this case, because the
+  condition *is* this repository's path. The test is therefore the only one that can fail
+  for an environmental reason, which is why it restores rather than only asserting.
+- The own copy is checked with the leak rules but not with the path rules. Naming
+  `src/ci/workflow.mjs` in this repository's own workflow is correct — it is the one place
+  that path resolves — so asserting against it would be asserting that this repository's
+  CI cannot mention its own source.
+- If this repository is ever vendored or installed as a package, `PACKAGE_ROOT` becomes a
+  directory under `node_modules` with no `.git` and no workflows, and the copy simply stops
+  existing there. Nothing needs to detect that case.
+
+## Rejected
+
+- **A `--managed` flag, or `init --refresh-ci`.** Rejected under ruling 1: it makes the
+  difference between "never overwrites your CI" and "will overwrite your CI" a value the
+  consumer sets. The ruling is worth more if there is no spelling for the wrong answer.
+- **Let `upgrade` refresh the file in every repo.** Rejected: it is the merge-strategy
+  problem ADR-023 already declined, and adding a repository to the exception list is
+  strictly smaller than solving it for all of them.
+- **One workflow file that detects which repo it is in at run time.** Rejected: a consumer's
+  file would then contain this repository's path and its commands, which is ADR-023 ruling 1
+  violated inside the very file ruling 1 was written for.
+- **Give the own copy a managed region marker instead of full replacement.** Rejected: this
+  repository has no hand edits to preserve, so a merge strategy would be machinery for a
+  case that does not exist here and would leak into the consumer copy's design.

simt_dev/vendored/docs/decisions/ADR-025-a-baseline-is-a-region-in-a-file-the-repo-owns.md

@@ -0,0 +1,153 @@
+# ADR-025 — a baseline is a region in a file the repo owns, and only a file that can hold one
+
+Status: accepted · 2026-10-06 · rulings 1 and 2 narrowed by ADR-026 · ruling 8 withdrawn by ADR-048
+**Affects:** `src/projection/baselines.mjs`, `src/projection/markers.mjs`,
+`src/projection/project.mjs`, `src/commands/init.mjs`, `test/baselines.test.mjs`,
+`test/leaks.mjs`, `test/surfaces.test.mjs`. Settles §3.1 row 5 and row 7 of
+`docs/plans/0001-centralise-agent-rules.md` for the files it names, and leaves the rest ruled
+out by name rather than outstanding.
+
+Rulings 1 and 2 did not survive whole: ADR-026 rules that a config whose grammar cannot
+carry a comment is owned whole rather than not at all, and does so by reusing this record's
+neighbour (ADR-023) rather than by overriding anything here. What rulings 1 and 2 still
+settle is narrower and is not in dispute — a file that **can** hold a region is a region,
+and no comment-capable file is to be taken over whole.
+
+## Context
+
+§2.2 of the plan named the mechanism before any of the files existed: *"Same for
+`.editorconfig`, `.prettierrc`, `.gitattributes`. Local content is never clobbered; shared
+content can never drift silently."* Only the first file was ever built, and only for
+`AGENTS.md`.
+
+§3.1 row 5 asks for eight baselines and promises six concrete resolutions: the `singleQuote`
+contradiction, the ten repos with no editorconfig, the missing strictness flags, the three
+disagreeing `.nvmrc`, no `packageManager`, and eslint fighting prettier in three of five
+configs. Row 7 asks `init` to scaffold them.
+
+What is actually missing is not the files. `rules/stacks/node.md` already mandates their
+shapes — *".nvmrc pins the development version"*, *"`moduleResolution` spelled consistently"*,
+*"Prettier owns formatting"* — and mandates them **with no values**. Every line is a shape
+with the value left to whoever reads it. That is the failure AGENTS.md names as "can the
+reader act on it — command, not location?", except it is worse: the reader is told a file
+exists and not what goes in it.
+
+The same profile is where the plan's own resolutions point. `singleQuote` is a
+`[prettier]`-section value, which prettier reads from `.editorconfig` — a file whose comment
+character is `#`, and therefore a file that can hold a region.
+
+**This paragraph is what ADR-048 withdraws.** A section header in editorconfig is a glob
+over paths, so `[prettier]` matches a file named `prettier`, and `singleQuote` is not one of
+the nine keys the specification defines — nor one of the six properties Prettier maps. The
+reasoning below is sound and the conclusion does not follow from it: the file *can* hold a
+region, and that is the whole of what makes it the right home for a shared claim. It says
+nothing about whether a given claim is one a reader honours. Ruling 8 is withdrawn for that
+reason and the record is kept, because the reasoning is what makes the withdrawal legible.
+
+## Rulings
+
+1. **A baseline is a marker-managed region, not a file simt-dev owns.** The repo keeps
+   `.editorconfig`; simt-dev owns the bytes between `# simt-dev:begin` and `# simt-dev:end`.
+   This is §2.2 as written. A whole-file baseline would overwrite the exact content the plan
+   says must never be clobbered, and the ten repos that already have an editorconfig are the
+   ones with the most to lose.
+
+2. **Only a file whose format can carry a comment may hold a region.** This is the whole
+   filter, and it is grammar rather than caution. `.editorconfig` qualifies: `#` is a comment
+   in INI and every editorconfig parser honours it. `tsconfig.base.json` and
+   `.prettierrc.json` do not — JSON has no comment syntax, and a "region" in a JSON file
+   would have to be a string literal that `tsc` and prettier both have to tolerate.
+   `.nvmrc` is a bare version string. A baseline for any of those is a whole-file write, and
+   ruling 1 forbids it. They are ruled out, not deferred.
+
+3. **The marker syntax is a property of the file, not of the module.** `markers.mjs` grew
+   `STYLES`: `html` for `AGENTS.md` and every `.mdc`, `hash` for `.editorconfig`. The HTML
+   rendering is byte-identical to what it was — asserted by the existing 36 tests — because a
+   change to those bytes would make every adopted repository drift. A third style is a row in
+   that table, not a second implementation, per the ADR-002 rule about a third copy being a
+   defect.
+
+4. **The region goes first, and `where` is a style property rather than a caller argument.**
+   editorconfig takes the last matching definition. An appended region would leave the repo's
+   `indent_size = 4` in the file and silently dead — no byte removed, nothing in `check` to
+   report, and a reader with no way to tell why their setting stopped working. Placing the
+   shared block first makes the repo win by construction. The placement belongs to the style
+   because the correct answer is a fact about the file's semantics, and a caller who has to
+   remember it is a caller who gets it wrong once.
+
+5. **A region that is not at the top is a finding of its own.** `check` raises `MISPLACED`
+   whenever bytes appear above `# simt-dev:begin`, with the reason and the fix. The bytes are
+   right and the meaning is wrong, which is the hardest drift to notice; a finding whose
+   detail said so is the only thing that makes it visible. `sync` repairs it by rewriting the
+   placement and leaving the content alone.
+
+6. **Absence blocks, the way absence of a rule file blocks.** An adopting repository ran
+   `init`; a baseline that has been deleted or a file whose region has been removed is drift,
+   and the remedy is one command. The alternative — silently reporting nothing for a file the
+   repo does not have — was rejected for the same reason ADR-023 ruling 4 exists: a check that
+   reports only what it can see is a check that cannot be relied on to report what it cannot.
+
+7. **`end_of_line` is deliberately absent.** It reads like an editor setting and is not one.
+   It decides whether a save rewrites every line of every file, which is a claim about what
+   belongs in the repository, and that claim is made in `.gitattributes` where git reads it.
+   Setting it here would change every checkout of every repository on the day it landed, and a
+   repository that wanted CRLF could only opt out by deleting the region — losing the
+   indentation baseline with it.
+
+8. **`[prettier] singleQuote = true` is in, `printWidth` is out.** `singleQuote` is the
+   contradiction the plan named first, it is the only comment-capable home prettier reads, and
+   a repository with its own prettier config overrides it — which is the right order, because
+   the nearer file wins. `printWidth` is absent because nobody has ruled on one. The
+   repository that already has `120` would be overruled by a number no one argued about, and a
+   baseline that picks a value the estate has not agreed to is not a baseline.
+
+   **The `singleQuote` half of this ruling is withdrawn by ADR-048**: `[prettier]` is a glob
+   that matches a file named `prettier`, and `singleQuote` is not a key the editorconfig
+   specification defines or that Prettier reads, so the block it ruled in applied to nothing.
+   The `printWidth` half stands and was never implemented — nothing has ruled on a number, and
+   that is unchanged.
+
+9. **`.gitattributes` is ruled out for now, for the reason in ruling 7.** It is
+   comment-capable and would qualify mechanically, but every line that earns its place in it
+   changes what git does to the working tree — `text=auto` rewrites every checkout,
+   `linguist-generated` and `export-ignore` need build directories, and build directories are
+   stack-specific, which §3.2 keeps local. A shared `.gitattributes` region today is either
+   dangerous or empty. It needs its own record, and this is not it.
+
+   **Ruling 9 is superseded by ADR-027**, which is the record this ruling asked for. The
+   deferral was sound and the conclusion was one line too broad: the shared region is neither
+   dangerous nor empty, it is exactly one line — `* text=auto eol=lf` — and the `auto` is
+   what makes it safe. The rest of this ruling stands, including ruling 7's refusal to
+   claim line endings from `.editorconfig`: they are claimed here instead, where a repository
+   can overrule the claim on one line without losing anything else.
+
+## Consequences
+
+- `.editorconfig` is the only baseline. `BASELINES` is a list, so the next one is a row, and
+  `test/baselines.test.mjs` asserts the list is exactly `['.editorconfig']` with a real style,
+  a reason and a record number on every row — a baseline added without noticing that its file
+  cannot hold a comment would fail there rather than in someone's editor.
+- The node profile's shapes now have values behind them for indentation. They still do not
+  for `singleQuote` — ADR-048 withdrew the claim that this file supplied one — nor for
+  `.nvmrc`, `packageManager` and `moduleResolution`, and the profile
+  still mandates the shapes. That gap is named in the plan's row 5 annotation rather than
+  hidden, because the next fix is three more JSON files and therefore a different mechanism.
+- `init`'s plan now prints a third caption, because a reader told "frontmatter kept, body
+  replaced" about a file that keeps everything outside a region has been told the wrong thing
+  about the edit they are approving.
+- This repository is not self-adopted and therefore has no `.editorconfig`. That is the same
+  reason it has no generated region in `AGENTS.md`.
+
+## Rejected
+
+- **Ship whole template files for every baseline in §3.1 row 5.** Rejected under ruling 1:
+  it overwrites repo content, and eight of the files in that row are JSON or bare strings,
+  which is the whole problem.
+- **A `tooling/` directory of template files, as §5 of the plan shows it.** Rejected for the
+  same reason: the directory was drawn before anyone checked which of those files can hold a
+  managed region, and four of the nine cannot.
+- **A manifest key to opt out of a baseline.** Rejected for now: with one universal baseline
+  there is nothing to opt out of, and a key added before there is a choice is a key nobody
+  reads. If a second baseline turns out to be stack-specific, the opt-out question returns with
+  it.
+- **`[prettier] printWidth = 100`.** Rejected under ruling 8.

simt_dev/vendored/docs/decisions/ADR-026-a-config-without-a-region-is-a-file-the-repo-owns.md

@@ -0,0 +1,117 @@
+# ADR-026 — a config whose grammar cannot hold a region is a file the repo owns
+
+Status: accepted · 2026-10-06
+**Affects:** `src/ci/dependabot.mjs`, `src/commands/ci.mjs`, `src/commands/init.mjs`,
+`src/commands/help.mjs`, `src/projection/project.mjs`, `test/ci.test.mjs`.
+**Narrows:** ADR-025 rulings 1 and 2. **Extends:** ADR-023 (bytes copied, written once),
+ADR-024 (simt-dev's own copies are refreshed by `simt-dev ci`). Settles the Dependabot
+item §3.1 row 8 of `docs/plans/0001-centralise-agent-rules.md` named as still missing.
+
+## Context
+
+ADR-025 ruled that a baseline is a marker-managed region, and that only a file whose format
+can carry a comment may hold one. Ruling 2 named `tsconfig.base.json`, `.prettierrc.json` and
+`.nvmrc` as "ruled out, not deferred", and its consequences promised that "the next fix is
+three more JSON files and therefore a different mechanism". That mechanism did not exist.
+
+Dependabot arrived first, and not by accident: it was the one owed baseline that needed no
+new mechanism at all, and the gap had been named and dated in the plan since before any of
+this existed — row 8 says the Dependabot config "is still missing and is not part of that
+ruling". A repo with no Dependabot config does not receive security patches as pull requests;
+it finds out from an advisory.
+
+Dependabot also supplied the missing counterexample to ADR-025's ruling 2, which is why
+this record exists rather than a template file. `updates:` is a **YAML list**. Two lists
+cannot be concatenated. There is no region shape that means "these entries are mine and
+those are yours", so there is nothing to merge and nothing to preserve — the mechanism
+ADR-025 chose does not apply to this file because the file's grammar has nowhere to put it.
+
+That is a property of the file, not a reason to prefer a different file. The question worth
+ruling on is what ownership means for a config simt-dev writes into a repo and must not
+rewrite, and ADR-023/024 had already answered it for the file sitting next to this one in
+the same directory.
+
+## Rulings
+
+1. **A config whose grammar cannot hold a region is a whole file the repo owns from the
+   moment it is written.** `simt-dev init` creates `.github/dependabot.yml` once and never
+   touches it again. There is no partial ownership to describe because there is nothing to
+   merge; the repo's own cadence, `ignore` rules, `groups` and private `registries` are
+   edits to a file simt-dev has already stopped writing.
+
+2. **Ownership does not depend on content being mergeable.** ADR-023's rule — copied bytes,
+   written once, the reader's from the first commit — extends from the CI gate to this file
+   unchanged. A reader of ADR-023 who assumed the rule depended on the file having a
+   place for a region was wrong, and this is the record that says so.
+
+3. **The content is detection-driven, from facts that already exist.** Each entry names an
+   ecosystem this tree has a manifest for, derived from `fingerprint`'s `has-*` facts and
+   stack tags. No new fact is declared, so ADR-013's "content is data, the engine is code"
+   holds without a profile change. The failure this prevents is an entry that looks
+   configured and configures nothing: Dependabot finds no manifest, opens no pull request,
+   and the config reads as though the repo's dependencies are watched.
+
+4. **One entry per ecosystem, not one per manifest.** `pyproject.toml` and `requirements.txt`
+   are two ways of declaring python dependencies and Dependabot reads both under `pip`. Two
+   `pip` entries is a duplicate the config rejects or silently collapses, and the reader has
+   no way to tell which won.
+
+5. **`has-build-system` is not mapped to an ecosystem.** Makefile, CMakeLists and
+   `meson.build` are build files, not dependency manifests, and Dependabot has no ecosystem
+   that reads them. Mapping the fact to `maven` would be a plausible-looking line that
+   watches nothing — ruling 3's failure mode, wearing a fact's name.
+
+6. **`github-actions` is the one unconditional entry,** and it is unconditional because of
+   what this package ships: simt-dev writes a workflow into every repo it adopts, so every
+   repo that adopted it has Actions to watch. A repo with no workflow has no Actions.
+
+7. **The file appears only where `.github/workflows` already exists,** for the same reason
+   ADR-023 ruled it for the gate: creating that directory silently would start opening
+   pull requests nobody asked for.
+
+8. **Both `.github` files are planned, written and refused together, by one command.**
+   `simt-dev ci` handles the workflow and the config because they answer the same question
+   the same way; a reader learning which command owns which file would be learning about
+   ownership twice to learn nothing new. `simt-dev ci` still refuses outside this
+   repository, naming `simt-dev init --yes` as the writer.
+
+9. **`check` reports absence and never contents,** as it does for the gate. A file whose
+   every byte is legitimately theirs cannot be diffed against a template without reporting
+   their work as drift.
+
+## Consequences
+
+`.github/dependabot.yml` joins `.github/workflows/simt-dev-check.yml` as a second shipped
+`.github` file with one ownership rule and one writer. `check` gains a second `absent`
+finding whose remedy is the same `simt-dev init --yes`, so a repo that runs Actions with a
+gate but no dependency updates is reported rather than left to find out from an advisory.
+
+What this does **not** settle: `tsconfig.base.json`, `.prettierrc.json` and `.nvmrc` are
+still regions that cannot exist. Ruling 1 answers the ownership question for them when
+someone writes one, and it is deliberately the same answer as here, but no file has been
+written and none should be inferred. The gap stays named in §3.1 row 5.
+
+The estate consequence is the one that costs: no repository in it had a Dependabot config,
+and the three pilot repositories from Phase 4 will each need `simt-dev upgrade` to receive
+one. That is ADR-001's ordinary cost and is not a defect.
+
+## Rejected
+
+- **A region inside `updates:`**, with entries above and below it. Rejected because it does
+  not exist: YAML lists have no interior, and a comment-delimited block in the middle of a
+  sequence is not a sequence. Shipping it would produce a config GitHub rejects.
+- **Merging by rewriting the whole `updates:` list,** preserving the repo's entries by
+  parsing and re-emitting them. This is the one design that could give partial ownership,
+  and it is rejected for the same reason as a manifest merge: simt-dev would have to parse
+  and re-emit a config it does not own, and every round trip through this tool would be an
+  opportunity to drop a key this tool never understood. ADR-015's content-not-location
+  principle cuts the same way — the repo's config is identified by its bytes, and this tool
+  does not hold those bytes.
+- **Prefixing the keys** to claim ownership, as ADR-022 does for rule frontmatter. Rejected
+  because these keys are GitHub's: Dependabot rejects `simt-package-ecosystem`. Ownership
+  here is expressed by not writing the file again, not by renaming what GitHub reads.
+- **One fixed list for every repo,** `github-actions` plus `npm`. Rejected by ruling 3: it
+  writes an `npm` entry into a repo with no `package.json`.
+- **A new `has-node-manifest` fact,** so the table reads uniformly. Rejected because `node`
+  is already a stack tag and `typescript` already a dependency name; a third name for one
+  observation is the duplication `fingerprint.mjs`'s own table header rules out.