feat(repo): repo budget counts .claude/rules

dokimi/changelog.d/20261030_180000_napetrov_budget-claude-rules.md

@@ -0,0 +1,15 @@
+### Added
+
+- **`repo budget` counts `.claude/rules`** — a `.claude/rules/**/*.md` file is a `rule` entry.
+  A rule in the root `.claude/rules/` whose `paths` frontmatter names no pattern is `start`; one
+  whose `paths` does, and every rule of a subdirectory's `.claude/rules/`, is `lazy`, unless a root `CLAUDE.md` or `AGENTS.md` imports it: then it is `start` whatever its `paths`. The
+  frontmatter is not counted, since Claude Code removes it before loading the rule, and an `@`
+  path in a rule is not followed as an import. A link to a directory under `.claude/rules` is a
+  warning: the rules behind it are not counted. A link under `.claude/rules` that does not resolve in
+  the tree is named too, whatever its suffix, since the tree cannot tell a dangling file from a directory; a
+  `CLAUDE.md` under `.claude/rules` is a rule, not guidance. Measured on Claude Code 2.1.304: a root rule
+  raised the first prompt by its size, `paths: []` included; a scoped or nested one loaded only
+  once the agent read a file it covers; a rule's `@extra.md` was not expanded. `start` was too
+  low by every root rule before; `repo_budget.v1` widens `kind` with `rule`. Rules are counted
+  for a harness profile with the new optional `guidance.claude_rules: true` (`claude-code`
+  only); the Codex and Copilot profiles leave it unset.

dokimi/docs/architecture.md

@@ -298,12 +298,16 @@ next release.
   non-empty body line, as Claude Code does. It warns on skills with no description or body text,
   skills with near-duplicate descriptions (identical descriptions, or Jaccard index >= 0.5
   after stop-word removal when both have 3+ words left), on AGENTS.md files turned off, on
-  unreadable files (non-UTF-8, dangling or out-of-tree link), and on imports past the harness's
-  hop limit, each of which also trips
-  `--fail-on-warning`. `start` is not an upper bound: it is too low by what it does not count
-  (`.claude/rules/*.md` files; only `@path.md` imports are followed, so `@notes.txt` adds nothing;
+  unreadable files (non-UTF-8, dangling or out-of-tree link), on imports past the harness's
+  hop limit, and on a link under `.claude/rules` to a directory or out of the tree (the rules behind
+  it are not counted), each of which also trips `--fail-on-warning`. For a profile with `guidance.claude_rules` (Claude Code),
+  `.claude/rules/**/*.md` files are `rule` entries: `start` in the root `.claude/rules/` when
+  `paths` names no pattern, `lazy` when it does or in a subdirectory's `.claude/rules/`, but `start` whatever its `paths` when a root CLAUDE.md or
+  AGENTS.md imports it, the frontmatter not counted. `start` is not an upper bound: it is too low by what it does not count
+  (only `@path.md` imports are followed, so `@notes.txt` adds nothing;
   skills are read only from real `.claude/skills/*/SKILL.md` directories at the root, so a skill
-  behind a linked directory is not seen, and a nested `<subdir>/.claude/skills/` skill, which Claude
+  behind a linked directory is not seen, nor are the rules behind one (a warning names it), and a
+  nested `<subdir>/.claude/skills/` skill, which Claude
   Code loads on the first read in that subdirectory, is not counted; and imports past the profile's `guidance.import_max_depth`,
   four hops for Claude Code, which warn naming what was skipped). It is too high by whatever the
   harness cuts from the skill listing: the listing is charged in full, and the harness cuts it past

dokimi/docs/repo-evaluation-contract.md

@@ -46,9 +46,13 @@ SPDX-License-Identifier: Apache-2.0
 > no LLM call): root CLAUDE.md files and their `@` imports plus the skill listing at
 > start (AGENTS.md files and their imports too, unless a root CLAUDE*, `.claude/CLAUDE.md` or
 > `CLAUDE.local.md` turns them off), nested guidance lazily, skill bodies on demand, and AGENTS.md
-> files a root CLAUDE* turns off. The `start` total is not an upper bound: it is too low by what it does
-> not count (`.claude/rules/*.md` files, non-`.md` imports, skills behind a linked directory, nested
-> `.claude/skills/` skills, and
+> files a root CLAUDE* turns off. For a profile with `guidance.claude_rules` (`claude-code`),
+> `.claude/rules/**/*.md` files are `rule` entries: at start in the root `.claude/rules/` unless
+> their frontmatter's `paths` names a pattern, lazily then and in a subdirectory's `.claude/rules/`; a rule a root CLAUDE.md or AGENTS.md
+> imports is `start` whatever its `paths`;
+> an `@` path in a rule is not followed. The `start` total is not an upper bound: it is too low by
+> what it does not count (non-`.md` imports, skills behind a linked directory, rules behind one,
+> which warns, nested `.claude/skills/` skills, and
 > imports past the profile's `guidance.import_max_depth`, four hops for Claude Code, which warns
 > naming what was skipped), and too high by whatever the harness cuts from the skill listing.
 > Usage:

dokimi/docs/what-it-measures.md

@@ -53,7 +53,7 @@ manifest it owns:
 
 - `dokimi repo scan` (L0, no LLM): do the links, paths, identifiers, commands
   and build labels the guidance names exist in the tree?
-- `dokimi repo budget` (no LLM): how many characters of guidance and skill
+- `dokimi repo budget` (no LLM): how many characters of guidance, `.claude/rules` and skill
   listing every model call pays for at start-up, what loads lazily or on
   demand, and which AGENTS.md files a CLAUDE.md turns off.
 - `dokimi repo reach` (L1): which guidance files does a given harness actually

dokimi/dokimi/commands/repo.py

@@ -343,7 +343,7 @@ def cmd_repo_budget(args: argparse.Namespace) -> None:
         print(f"✗ {exc}", file=sys.stderr)
         sys.exit(EXIT_USAGE)
     _, tree = _source_tree(args)
-    payload = budget(tree, profile.guidance_names, profile.import_max_depth)
+    payload = budget(tree, profile.guidance_names, profile.import_max_depth, profile.reads_claude_rules)
     if args.out:
         try:
             save_artifact("repo_budget", payload, Path(args.out))