Address code-review findings on the init.go split

- Normalize os.WriteFile's mode literal to 0o644 in writeInitConfig,
  matching the sibling call in writeScaffolds in the same file.
- Document the archive-sharding convention this cycle improvised
  (split into a numbered shard, never bump max-file-length without
  consent) in the audit-checklist so future audits don't reinvent it.
- Fold setInitUsage's identical missing-unit-test gap (surfaced by
  this review, not the original audit sweep) into plan/2607191917
  alongside printInitCatalog's.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017Req4a3SADus6KRYrmYqKg

Author: claude

PLAN.md

@@ -253,6 +253,6 @@ footer: |
 | 2607082052 | ✅     | sonnet | [SARIF output format for `mdsmith check`](plan/2607082052_check-sarif-output.md)                                                                        |
 | 2607121915 | ✅     | sonnet | [Resolve the internal/linkgraph purity-contract mismatch for wikilink resolution](plan/2607121915_arch-fix-linkgraph-wikilink-purity.md)                |
 | 2607170527 | 🔳     | opus   | [External link checking on WASM hosts (MDS072)](plan/2607170527_wasm-external-link-check.md)                                                            |
-| 2607191917 | 🔲     | haiku  | [Add a dedicated unit test for printInitCatalog](plan/2607191917_arch-fix-printinitcatalog-unit-test.md)                                                |
+| 2607191917 | 🔲     | haiku  | [Add dedicated unit tests for printInitCatalog and setInitUsage](plan/2607191917_arch-fix-printinitcatalog-unit-test.md)                                |
 | 2607191918 | 🔲     | haiku  | [Deduplicate isClaimed between internal/schema and requiredstructure](plan/2607191918_arch-fix-isclaimed-dedup.md)                                      |
 <?/catalog?>

cmd/mdsmith/init.go

@@ -140,7 +140,7 @@ func writeInitConfig(configFile, fromMarkdownlint, starterName string, force boo
 	if err != nil {
 		return err
 	}
-	if err := os.WriteFile(configFile, data, 0644); err != nil {
+	if err := os.WriteFile(configFile, data, 0o644); err != nil {
 		return fmt.Errorf("writing %s: %w", configFile, err)
 	}
 	if source != "" {

docs/development/architecture-audit.md

@@ -65,14 +65,17 @@ None.
   matching `init_unit_test.go`. `main.go` is now 678 lines.
   No behavior change; `go test ./...` and
   `go tool golangci-lint run` are green.
-- `printInitCatalog` (`cmd/mdsmith/init.go`) has no
-  dedicated unit test — only the e2e `mdsmith init --list`
-  subprocess test exercises it.
+- `printInitCatalog` and `setInitUsage`
+  (`cmd/mdsmith/init.go`) have no dedicated unit test —
+  only e2e subprocess tests (`mdsmith init --list` and
+  `--help`) exercise them.
   [tests.md][tests]: "A new function lands together with
   its dedicated unit test by name," and an e2e test
   reachable without the process boundary is an inverted
-  pyramid. Not a public surface itself (a `runInit`
-  helper), so tax not blocker —
+  pyramid. Neither is a public surface itself (both are
+  `runInit` helpers), so tax not blocker. The
+  `setInitUsage` half of this surfaced during this cycle's
+  3x code-review pass on the fix, not the original sweep —
   [plan/2607191917][2607191917].
 - `internal/rules/requiredstructure/rule.go`'s `isClaimed`
   is a byte-for-byte copy of

docs/development/architecture/audit-checklist.md

@@ -59,6 +59,25 @@ budget.
   per-finding "suggested fix". Single
   sentences over ~30 words trip MDS023 /
   MDS024 on the audit log.
+- **When the archive itself crosses the
+  300-line budget**: do not bump
+  `max-file-length` on it — that needs
+  explicit user consent per
+  [CLAUDE.md](../../../CLAUDE.md)'s
+  "Never modify `.mdsmith.yml` ...
+  without explicit user consent" rule.
+  Instead, split it the same way the
+  main log split into it: move the
+  oldest entries into a new numbered
+  shard (`architecture-audit-archive-2.md`,
+  then `-3.md`, ...), moved verbatim, not
+  reworded. Each shard links to the next
+  one it overflowed into, and the
+  original `architecture-audit-archive.md`
+  stays the entry point new readers find
+  first. Re-run `mdsmith fix .` after the
+  split so every parent catalog picks up
+  the new shard.
 
 ## Initial file template
 

plan/2607191917_arch-fix-printinitcatalog-unit-test.md

@@ -1,22 +1,23 @@
 ---
 id: 2607191917
 title: >-
-  Add a dedicated unit test for printInitCatalog
+  Add dedicated unit tests for printInitCatalog and
+  setInitUsage
 status: "🔲"
 model: haiku
 summary: >-
-  printInitCatalog (cmd/mdsmith/init.go) has no
-  TestPrintInitCatalog; only an e2e subprocess test
-  exercises `mdsmith init --list`. Flagged by the
-  2026-07-19 audit as an inverted-pyramid tax finding.
+  printInitCatalog and setInitUsage (cmd/mdsmith/init.go)
+  have no dedicated unit test; only e2e subprocess tests
+  exercise them. Flagged by the 2026-07-19 audit and its
+  own 3x code-review pass as inverted-pyramid tax findings.
 ---
-# Add a dedicated unit test for printInitCatalog
+# Add dedicated unit tests for printInitCatalog and setInitUsage
 
 ## Goal
 
-Give `printInitCatalog` its own unit test so `mdsmith init
---list` behavior is locked at the unit layer, not only
-through the e2e subprocess test.
+Give `printInitCatalog` and `setInitUsage` their own unit
+tests. Lock `mdsmith init --list` and `--help` behavior at
+the unit layer, not only through e2e subprocess tests.
 
 ## Background
 
@@ -39,9 +40,22 @@ found this gap:
   writes to an `io.Writer`. It needs no subprocess to
   exercise.
 
-This is `tax`, not `blocker`. `printInitCatalog` is a
-helper `runInit` calls. It is not itself a CLI subcommand
-entry point, an LSP handler, or a `rule.Rule` method.
+A follow-up 3x code-review pass on the PR that filed this
+plan found the identical shape one function over:
+
+- `setInitUsage` in [init.go](../cmd/mdsmith/init.go) has
+  no `TestSetInitUsage` anywhere in
+  [init_unit_test.go](../cmd/mdsmith/init_unit_test.go).
+- Its only exercise is the e2e subprocess assertion for
+  `mdsmith init --help` in
+  [e2e_coverage_test.go](../cmd/mdsmith/e2e_coverage_test.go).
+- Same tests.md citation applies: `setInitUsage` installs a
+  closure on a `*pflag.FlagSet` and needs no subprocess to
+  exercise directly.
+
+Both are `tax`, not `blocker`. Neither is itself a CLI
+subcommand entry point, an LSP handler, or a `rule.Rule`
+method — both are helpers `runInit` wires up.
 
 ## Tasks
 
@@ -51,16 +65,28 @@ entry point, an LSP handler, or a `rule.Rule` method.
    ("Starters (mdsmith init --starter <name>):" and "Packs
    (mdsmith init --add <name>):") plus at least one known
    starter name and one known pack name appear in the output.
-2. Leave the existing e2e `--list` test in place — it still
-   covers the full CLI dispatch path — but do not duplicate the
-   catalog-content assertions there beyond a smoke check.
-3. `go build ./...` passes.
-4. `go test ./cmd/mdsmith/...` passes.
+2. Add `TestSetInitUsage` in
+   [init_unit_test.go](../cmd/mdsmith/init_unit_test.go),
+   building a `*pflag.FlagSet`, calling `setInitUsage`, then
+   invoking `fs.Usage()` with a captured `os.Stderr` (or by
+   redirecting the closure's writer, whichever the existing
+   test helpers in this file support) and asserting the
+   printed text names `--starter`, `--from-markdownlint`,
+   `--add`, `--force`, and `--list`.
+3. Leave the existing e2e `--list` and `--help` tests in
+   place — they still cover the full CLI dispatch path — but
+   do not duplicate the content assertions there beyond a
+   smoke check.
+4. `go build ./...` passes.
+5. `go test ./cmd/mdsmith/...` passes.
 
 ## Acceptance Criteria
 
 - [ ] `TestPrintInitCatalog` exists in
       [init_unit_test.go](../cmd/mdsmith/init_unit_test.go) and
       exercises `printInitCatalog` directly (no subprocess).
+- [ ] `TestSetInitUsage` exists in
+      [init_unit_test.go](../cmd/mdsmith/init_unit_test.go) and
+      exercises `setInitUsage` directly (no subprocess).
 - [ ] `go test ./...` is green.
 - [ ] `mdsmith check .` is green.