G17-G22: OSIL now reads C + a correctness-gated loop optimizer

[cemphlvn]

Adds a C lowering profile, a mechanical loop optimizer, a capability model that prices its own blind spots, and a gate for the harness discipline that turned out to be the dominant failure mode.

Gates closed:
G17 C as an ecosystem profile with a tested preservation contract
(ADR-0014). may_lose { declared_licence } is the repo's first
may_lose field backed by an experiment rather than inspection:
restrict on every array pointer across 151 TSVC loops recovers 0.
G19 a lifter recovers dependence structure from C source alone (libclang),
reproducing a hand analysis measurement had already confirmed, 10/10.
G20 a chooser turns those facts into a transformation under three gates --
legal, correct, faster -- across four families.
G21 the architecture analyses its own reach: capabilities are declared
vocabulary (grammar v0.7), and the ceiling on what any chooser could
recover is DERIVED from them. just price prices a capability before
it is built.
G22 the harness discipline is checked rather than remembered. C-Reduce
named this bug class in 2012; it appeared eight times here.

Results, stated with their limits:
clang -O3 alone 64/151 = 42.4%
clang + pipeline 74/151 = 49.0% 10 kernels, all bit-identical or 1e-6
Every transformation is textbook 1980s work. No tuned baseline was beaten.
The published record (GCC 56.0%) is on different hardware and is NOT
comparable. The analyser's own ceiling is 52.3% -- below the record, and
knowable before the attempt started.

What the record shows going wrong, kept deliberately:

• a pre-registered kill condition fired and was honoured (record-attempt.md)
• s221 is RETRACTED as evidence for the speed gate: 1.0x under background load, 1.27x quiet. Contention changes speedups up to 3.8x and flips accept/reject (measurement-contention.md)
• three wrong-code emissions caught by the correctness gate, including a scalar-dependence unsoundness that would have compiled and run
• a false retraction of the Isaria citation, corrected from the primary PDF
• a supermodularity claim about Wall 1993 corrected against the paper

demo/ is a standalone 30-second repro: one TSVC loop that clang and gcc both refuse to vectorize under -O3, -ffast-math, -Ofast, and both suggested pragmas -- next to a bit-identical rewrite worth ~5x. It asserts nothing; it runs the reader's own compilers and prints what they say.