Make bisect_ppx actually instrument the library — coverage was measuring nothing (#713)

The coverage job runs, passes, and **measures nothing**.

```console
$ dune runtest --force --instrument-with bisect_ppx
537 tests OK, exit 0
$ bisect-ppx-report summary
Error: no *.coverage files found
```

Zero `.coverage` files are written anywhere in the tree.

## Root cause

`lib/dune` declares `(preprocess (pps …))` but **no instrumentation
backend**. `--instrument-with bisect_ppx` only instruments libraries
that opt in with an `(instrumentation (backend bisect_ppx))` stanza.
Without one, dune instruments nothing, the tests run uninstrumented, and
there is no coverage data to report.

## Why nobody noticed

`.github/workflows/ci.yml:213` sets **`continue-on-error: true`** on the
step that calls `bisect-ppx-report`. So the report failed on every run,
the failure was swallowed, and the coverage job went green while
producing no data — the same hollow-gate shape as a test that passes
without running.

The `continue-on-error` is **left in place**: it's defensible for a
visibility-only artefact step, and with this fix the step now has
something to report. Worth revisiting separately if coverage ever
becomes gating.

## After the fix

| | before | after |
|---|---|---|
| `.coverage` files | **0** | 2 |
| Coverage | *unmeasurable* | **8103/16243 (49.89%)** |

That number has never been measurable before, so treat it as a
**baseline** — not a regression, and not an improvement.

## What it says needs attention

Core paths are reasonably covered:

| Module | Coverage | |
|---|---|---|
| `parser.ml` | 75.64% | 969/1281 |
| `quantity.ml` | 79.28% | 241/304 |
| `typecheck.ml` | 68.37% | 1042/1524 |
| `borrow.ml` | 68.03% | 800/1176 |
| `resolve.ml` | 59.21% | 405/684 |
| `lexer.ml` | 58.33% | 49/84 |

The gaps are concentrated in **backends and tooling**, not the front
end:

| Module | Coverage | |
|---|---|---|
| **`wasm_gc.ml`** | **0.00%** | **0/178** ← entirely unexercised |
| `wasm.ml` | 0.24% | 1/423 |
| `types.ml` | 9.95% | 19/191 |
| `lean_codegen.ml` | 10.48% | 13/124 |
| `value.ml` | 13.13% | 26/198 |
| `c_codegen.ml` | 14.60% | 67/459 |
| `lsp_server.ml` | 15.47% | 43/278 |

`wasm_gc.ml` at **0/178** is the standout — 178 points, none reached by
any of the 537 tests.

## Two things I checked and did *not* change

**The 32 `[SKIP]`s in `res-to-affine-walker` are correct behaviour, not
a fake gate.** `test_walker.ml:65` skips with an actionable message when
the tree-sitter CLI isn't on PATH, and CI *does* install it
(ci.yml:59–68) — so those tests run in CI and skip only locally. Good
design; left alone.

**The suite is otherwise healthy:** 537 tests run, 0 fail. Better than
the estate norm.

## Scope

One file changed: `lib/dune`.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Author: hyperpolymath

lib/dune

@@ -114,7 +114,15 @@
  (flags
   (:standard -w -8-9))
  (preprocess
-  (pps ppx_deriving.show ppx_deriving.eq ppx_deriving.ord sedlex.ppx)))
+  (pps ppx_deriving.show ppx_deriving.eq ppx_deriving.ord sedlex.ppx))
+ ; Without this stanza `dune runtest --instrument-with bisect_ppx` has nothing
+ ; to attach to: dune instruments nothing, the tests run uninstrumented, and no
+ ; .coverage files are written. `bisect-ppx-report summary` then fails with
+ ; "no *.coverage files found" — and because the CI step that calls it carries
+ ; `continue-on-error: true`, that failure was swallowed and the coverage job
+ ; went green while measuring nothing.
+ (instrumentation
+  (backend bisect_ppx)))
 
 (menhir
  (modules parser)