diff --git a/CLAUDE.md b/CLAUDE.md
index c5629db..33aa4f1 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -187,6 +187,60 @@ Key invariants in the code:
   `git commit` is preceded by an explicit message proposal
   and user approval. No AI attribution in commit messages.
 
+## Build hygiene
+
+`target/` is git-ignored and 100% regenerable, but it grows
+without bound — cargo never garbage-collects old hash-suffixed
+artifacts, so stale test binaries (each ~100 MB, statically
+linking the bundled engine + debug info), incremental-compile
+caches, and orphaned example binaries pile up across sessions
+(it reached **~38 GB** before the first sweep).
+
+Two prevention levers are configured in `Cargo.toml` `[profile.dev]`
+(the `test` profile inherits both):
+
+- **`incremental = false`** — the incremental cache alone reached
+  **16 GB** here (≈28 compilation units × every historical config,
+  never evicted), for little benefit in a full-`cargo test` workflow.
+  Off, it never regrows; the cost is whole-crate recompiles instead of
+  partial — seconds for a crate this size.
+- **`debug = "line-tables-only"`** — the default `debug = 2` is
+  ~85–90 % of each test binary; line tables keep file:line in panics
+  and backtraces (we debug via `tracing` logs) at a fraction of the
+  size.
+
+Even with those, stale artifacts still accumulate (cargo has no
+target/ eviction). **Run `cargo sweep` every now and then** to reclaim
+that — `cargo-sweep` (installed) prunes everything *except* the current
+build's artifacts:
+
+- **Keep only the current build (the usual sweep):** stamp,
+  build, then delete everything the build didn't touch —
+  ```
+  cargo sweep --stamp
+  cargo build --all-targets        # touch the artifacts to keep
+  cargo sweep --file               # remove everything older than the stamp
+  ```
+  Add `--dry-run` to `--file` first to preview what goes. Caveat:
+  `build --all-targets` only updates the mtime of what it
+  actually (re)builds, so already-fresh *dependency* artifacts
+  fall before the stamp and get swept too — they recompile once
+  on the next build (a one-time cost; everything is regenerable).
+  The first 38 GB → 20 GB sweep freed **19 GiB** this way.
+- **Lighter routine options:** `cargo sweep --time 30` drops
+  artifacts untouched for 30+ days; `cargo sweep --maxsize 10GB`
+  trims oldest until under a size cap.
+- **`--installed` is *not* the tool for same-toolchain cruft.**
+  It keeps only artifacts from currently-installed rustup
+  toolchains, so it frees space *only after you uninstall/replace
+  a toolchain*. For the usual "many builds, one toolchain"
+  accumulation it cleans **nothing** (verified: it would free 0
+  here) — use the stamp/file workflow instead.
+
+A good cadence is a sweep between major milestones (e.g. at
+session handoff). `cargo clean` remains the nuclear option (wipes
+all of `target/`, forcing a full from-scratch rebuild).
+
 ## Things deliberately deferred
 
 These are explicitly tracked (mostly in `requirements.md`) but
@@ -202,11 +256,11 @@ not yet implemented:
 - **m:n convenience** (C4): auto-generates a junction table
   with appropriate FKs — depends on relationships being solid
   (they are).
-- **Friendly error layer** (H1): partial — FK errors are
-  enriched both ways; full SQL→English translation pending.
 - **Strong syntax-help in parse errors** (H1a): point users at
   missing keywords/clauses rather than the unexpected
-  character.
+  character. *(H1 — the friendly **database**-error layer — is
+  done, ADR-0019; H1a is its separate parse-error sibling,
+  ADR-0021, still partial.)*
 - **Tutorial/lesson system**: acknowledged as in scope for
   design; needs its own ADR.
 - **Session log + Markdown export** (V4): the bigger UX
diff --git a/Cargo.toml b/Cargo.toml
index a42d918..b11af84 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -45,6 +45,21 @@ insta = { version = "1.47.2", features = ["yaml"] }
 pretty_assertions = "1.4.1"
 tempfile = "3.27.0"
 
+# Dev/test build hygiene (see CLAUDE.md "Build hygiene"). `cargo test`
+# links ~25 separate integration-test binaries, each statically
+# embedding the bundled engine + every dependency; the `test` profile
+# inherits these from `dev`.
+# - incremental = false: the incremental cache reached 16 GB here for
+#   little benefit in a full-suite workflow; disabling it keeps target/
+#   bounded.
+# - debug = "line-tables-only": full debug info (the default `debug = 2`)
+#   is ~85-90% of each test binary; line tables keep file:line in panics
+#   and backtraces (we debug via `tracing` logs, not a step-debugger) at
+#   a fraction of the size.
+[profile.dev]
+incremental = false
+debug = "line-tables-only"
+
 [lints.rust]
 unsafe_code = "forbid"
 unreachable_pub = "warn"