# Plan: consolidate integration tests into one binary

**Date:** 2026-06-02 · **Status:** done

## Outcome (measured)

| Metric | Before | After |
|---|---|---|
| `target/debug/deps` | 1.5 GB | **629 MB** (−58%) |
| Integration-test binaries | 26 | **2** (`it` + `typing_surface_matrix`) |
| Clean test build | 22.9 s | 22.4 s |
| Lib-edit relink | 13.3 s | 12.4 s |
| Tests | 2151/0/1 | 2151/0/1 (identical) |

cargo auto-detected `tests/it/main.rs` as the `it` target — no
`[[test]]` entry, no fixups, no warnings.

**Correction to the build-time hypothesis:** the expected relink
speed-up (≈13 s → ≈5 s) did *not* materialise. Wall-clock is dominated
by *compiling* (the lib + ~290 deps + the test code's
codegen/monomorphisation), not by *linking*. Consolidation collapses 26
link steps into 2 — which is why **disk** drops 58 % (each link emits a
whole standalone binary) — but compile work is unchanged, so build time
barely moves. **The real, durable win is disk, not speed.** Minor cost:
editing a single test file now recompiles the whole `it` module set
(test-only edits marginally slower; lib edits marginally faster — a
wash). Kept for the disk benefit, which is exactly the concern that
prompted it.

## Why

Each top-level `tests/*.rs` is a separate integration-test *crate* →
its own binary, each statically linking the bundled engine + every
dependency + debug info. There are **26** of them. With incremental
disabled (build-hygiene change, commit `42f9553`), an edit to the lib
forces **all 26 to relink**.

**Measured baseline:**
- Lib-edit relink (touch `src/app.rs`, `cargo test --no-run`): **13.3 s**
- Clean build of all test targets (`cargo clean && cargo test --no-run`): **22.9 s**
- `target/debug/deps`: **1.5 GB**

Collapsing to one binary turns 26 links into 1 — the dominant cost in
both numbers.

## Approach

The pattern already used by `tests/typing_surface_matrix.rs` (one
binary that `mod`-includes the `tests/typing_surface/` modules). Apply
it to the other 25 standalone files.

Feasibility pre-checks (done, all clear):
- **No top-level integration test uses insta** → moving files causes
  no snapshot-path churn.
- **No crate-level `#![…]` attributes** to hoist.
- **No shared `mod common;` helpers** to relocate; the 25 files are
  self-contained (`use rdbms_playground::…`), no inter-file `mod`.

## Steps

1. `mkdir tests/it`.
2. `git mv` the 25 standalone `tests/*.rs` → `tests/it/<name>.rs`
   (everything except `typing_surface_matrix.rs`, which stays as its
   own already-consolidated binary).
3. Create `tests/it/main.rs` = `mod <name>;` for all 25.
4. Confirm cargo builds it as one target (`tests/it/main.rs`
   auto-detection; add an explicit `[[test]] name="it"` only if not).
5. Fix anything the combined build surfaces (expected minimal): a
   stray `use crate::` self-ref, or a clippy `dead_code`/unused-import
   the per-crate split had hidden.

**End state: 2 integration-test binaries (`it` + `typing_surface_matrix`)
instead of 26.**

## Verification / go-no-go gate

- Full `cargo test` → identical **2151 / 0 / 1**.
- `cargo clippy --all-targets -- -D warnings` clean.
- Re-measure relink + clean build; record before/after in the commit.
- If the after-numbers aren't materially better, **revert** (all
  `git mv`, trivially reversible).

## Tradeoff (accepted)

Lose per-crate isolation: a compile error fails the whole `it` binary,
and the 25 files share one crate namespace (module-scoped — no name
clashes). Negligible for a solo teaching project; the build-speed win
is the priority.