[DanielLuning]

Merge #70 first. This branch builds on defk-o08.4/stmt-reads-once, because both add entries to docs/performance.md. Until #70 merges, this PR shows #70's commit too.

What changed

For each kept bounds check, the fast-path report searched for a small fact that would drop it: the "(it would follow from w >= 0)" hint, from Facts.residual/5. The search took about 2.4% of compile time, and 6% on SatEx.Kernels, in every compile, whether or not anyone read the report.

• The hint search runs only under use Defk, report: :full (the maintainer's choice). Defk.__before_compile__ passes hints: to Defk.Compiler.compile/4, which keeps it in the ctx. Without the option, a kept note ends at what nothing proves: x[i - w]: bounds check kept; nothing proves i - w < length(x).
• Tests: the two test modules that read hints (condition_facts_test, product_facts_test) use report: :full. A new module, Defk.ConditionFactsTest.Plain, checks that the same kernel without the option keeps the plain note.
• Docs: the report: :full option in the Defk moduledoc, docs/fast-path-report.md (the section is now "Hints and vectorisation"), docs/language.md (two places), and the Unreleased CHANGELOG.md entry that introduced hints. Hints have not shipped in a release, so that entry is corrected instead of adding a "Changed" line.
• The gate tool: DEFK_DUMP_C now also writes <Module>.report, each module's report notes, beside <Module>.c. The same diff -r then checks the report too. docs/agents/runs.md describes it.
• docs/performance.md 6.41: the result, and the two cuts tried first that keep every report the same, with the reasons they were not adopted:
  • A hint rendered only when its note survives saves 1.5% of the work but costs about 9% more time, because each kept note's closure keeps the env and the facts scope alive until the compile ends.
  • @hint_budget 8 already drops hints, and saves only 0.5%.

The branch has two commits of its own. The first records the two failed cuts and adds the report dump. The second makes the change.

Results

Measured with the DEFK_COMPILE_TIME harness, two interleaved rounds against #70's state:

• Reductions: 114.9 M → 110.8 M (−3.6%; −1.2% without SatEx.Kernels, which alone saves 3.2 M).
• Time: 932 → 922 ms (−1.1%, within drift).
• The budget: against the compiler before the symbolic-bounds chain, the work is now +3.5% (−2.2% without SatEx.Kernels).
• The C: unchanged. The two test modules that took report: :full differ only in their path-hashed defk:loop markers.
• The reports: the same, less their hints. Two notes of HashChain.Kernels that differed only by their hints (two inline sites) are now kept once, as the report keeps any duplicate note once. PCM.Kernels keeps its hints, because it uses report: :full.

Provisional: the laptop was on battery, with no quiet-machine check.

How it was checked

• mix ci on the laptop (macOS) passed: 2304 tests, Credo found no issues, Dialyzer reported 0 errors, every generated doc region up to date.

• The gate: the C diff, and the report diff with the hints removed from the baseline.

• Touches a promised surface? Yes: the public report: :full option now also turns on the hints. The Unreleased CHANGELOG.md entry that introduced hints says so. No deprecation is needed, because nothing is renamed or removed from a release.

• Touches test/stability/? No.