tests: CONTEXT_SLOTS index allocation is by-convention and unenforced -- input.rs's range comment is already wrong and it cost a flaky gate

[melihucar]

Gate: default
Agent: codex

Every #[cfg(test)] mod tests under crates/frankenrust-core/src/
compiles into one lib test binary, and they all share the
process-global CONTEXT_SLOTS table (context.rs's pub static CONTEXT_SLOTS). Two tests that pick the same thread_index therefore
race over one slot.

Nothing enforces that they don't. The only defence is a hand-maintained
prose comment above each file's index block, saying which ranges the other
files use. That is five comments that must be kept in sync by hand, and at
least one of them is already wrong.

The concrete wrongness today

callbacks/input.rs:274-277 says its range is 120-127, and its named
IDX_* block (input.rs:278-285) does stop at 127. But four more indices
are declared as function-local const IDX inside test bodies:

• input.rs:408 -> 128 (go_read_cookies_percent_decodes_...)
• input.rs:422 -> 129
• input.rs:434 -> 130 (go_read_cookies_with_no_cookie_header_returns_null)
• input.rs:446 -> 131 (go_read_cookies_with_a_single_empty_cookie_value_returns_null)

So input.rs's real range is 120-131, and neither its own comment nor
output.rs:410-413's cross-reference says so.

What it already cost

#169 added tests that claimed 130 and 131 on the strength of that comment.
The result was an intermittent red gate: input.rs's set(130, ..)
replaces the worker test's context and drops its counting
CompletionSignal unfired, so
finish_php_request_closes_the_context_fires_the_signal_once_and_leaves_the_slot_installed
fails its fire_count == 1 assertion. Measured at 1 failure in 300 runs
of the filtered set at --test-threads=8 (two reviewers independently
measured 10/400 and 2/60). The full suite passed 300/300, which is why the
gate went green -- the margin is an accident of the current test count.

That collision is fixed in #169's own file (moved to 132-134). The
mechanism is not fixed, and the next agent to add a slot-using test
reads the same wrong comment.

Note the quieter failure direction too: when the replacement context is
the one installed, input.rs's two tests still assert is_null() and
pass vacuously. A collision can hide a real regression as easily as it
can invent a fake one.

What to build

Replace the convention with a single enforced allocation point.

1. Add one #[cfg(test)]-only module -- suggest pub(crate) mod test_slots in crates/frankenrust-core/src/context.rs, next to
   CONTEXT_SLOTS itself, or a new src/test_slots.rs gated on
   #[cfg(test)] -- holding one named pub(crate) const per test that
   needs a slot, for all of misc.rs (1-4), servervars.rs (60-76),
   output.rs (100-116), input.rs (120-131) and callbacks/worker.rs
   (132-134).
2. In that module, list every constant once more in a
   const ALL: &[(&str, usize)] and add a #[test] that asserts the
   values are pairwise distinct, naming both colliding constants in the
   failure message. This is the part that makes it permanent: a future
   collision becomes a deterministic red test with a clear message instead
   of a sub-1% flake in an unrelated file.
3. Repoint all five test modules at test_slots::* and delete the
   per-file range comments, including the four function-local const IDX
   declarations in input.rs, which are the ones the comments missed.
4. Keep the indices small and dense. ContextSlots::slot grows the
   table one entry per index up to the one asked for and never reclaims
   (see core: ContextSlots::slot grows the slot table densely and unboundedly, and a wrong thread index costs 355 MB with a green gate #112, and the measurement in misc.rs's tests module comment:
   ~800k costs 362 MB of peak RSS). Reuse the existing 1-134 numbering
   rather than renumbering into tidy per-file blocks with gaps.

Acceptance

• No usize slot-index literal remains inside any
  crates/frankenrust-core/src/callbacks/*.rs test module; they all name
  a test_slots constant.
• The uniqueness test exists and fails (with both names in the message)
  when two constants are given the same value. Verify by temporarily
  duplicating a value; do not commit that.
• rg -n 'usize = [0-9]+' crates/frankenrust-core/src shows slot indices
  only in the new module.
• bash scripts/gate.sh default passes.

Out of scope

• core: ContextSlots::slot grows the slot table densely and unboundedly, and a wrong thread index costs 355 MB with a green gate #112's actual fix (making ContextSlots::slot sparse or bounded). This
  issue only keeps the test indices small; it does not change
  ContextSlots.
• Any production code. This is test-support only.
• context.rs's own tests, which build a local ContextSlots::new() and
  never touch the global table.