research: add observer-driven scale compiler S0–S2 (NARROW)

[mountain]

Outcome

This draft PR publishes S0--S2 for #140 with an intentionally narrow
disposition:

NARROW — a bounded raw-special-function-to-local-chart certificate
pipeline, not a general asymptotic compiler or a uniform special-function
evaluator.

The S0/S1 freeze remains unchanged. S2 was publicly frozen before
implementation at commit db9898888402d7be8bbd7458c7e6b7d86d011497 and
has a separate post-freeze manifest.

Earned results

• exact descent criterion (\ker q\subseteq\ker(q'\circ U)) and
  information-minimal joint-image repair;
• delayed-truncation witness for ((1+N^{-1})^N);
• unhinted polynomial Airy chart
  (t=N^{-1/3}u,\ z=N^{-2/3}\xi);
• strict held-out cusp chart, with no frozen-file changes:
  (t=N^{-1/4}u,\ p=N^{-1/2}\xi,\ q=N^{-3/4}\eta);
• analytic-germ/Newton exposed-face criterion and the finite-jet condition
  ((K+1)w_{\min}>\kappa);
• raw registered input besselj(N, N*z) lowers through a versioned DLMF
  representation bridge, then the generic germ path derives
  (\theta=N^{-1/3}\hat\theta) and (z-1=N^{-2/3}\hat\delta);
• machine-visible claim split:

local_chart_certified = true
uniform_integral_certified = false

• typed refusals for noninteger-domain use of the integer Bessel registry,
  unregistered bessely, unsafe sin(N*x) tail semantics, regular
  Gaussian saddles, and competing balances;
• independent black-box replay passes the raw Bessel, non-Bessel cubic,
  quartic cusp, domain, and oracle-leak controls.

Why the claim remains NARROW

The raw bridge selects one classical representation from a registry; it does
not discover integral representations. The local certificate does not prove
contour localization, remote-tail cancellation, Stokes/branch control, a
uniform error bound, or net economy.

SymPy 1.14 does not complete the same raw coupled-scale task, but can verify
the germ after representation information is supplied. No executable Wolfram
kernel or authenticated cloud path was available, so official capability
documentation is not counted as comparative execution evidence.

Surreal numbers remain semantically useful for organizing larger scale
ensembles, but they are not required by this exact rational Newton/germ
runtime. No surreal implementation is added.

Validation

(cd sonnet/observer-driven-scale-chart-compiler/prototype &&
  python -m unittest discover -s tests -v)

(cd sonnet/observer-driven-scale-chart-compiler/analytic-germ &&
  python -m unittest discover -s tests -v &&
  sha256sum -c POST_FREEZE_MANIFEST.sha256)

python sonnet/observer-driven-scale-chart-compiler/math/verify_bessel_certificate.py
python sonnet/observer-driven-scale-chart-compiler/benchmarks/run_s2_raw_baselines.py
python sonnet/observer-driven-scale-chart-compiler/benchmarks/run_s2_compiler_red_team.py

Observed:

frozen S0/S1 tests: 12/12
post-freeze S2 tests: 15/15
post-freeze manifest: 10/10
mathematical replay: pass
independent compiler red team: pass
raw SymPy baseline: no coupled chart discovered
Wolfram exact row: not executed

Architecture boundary

The changing symbolic grammar, representation registry, domain/branch
assumptions, and SymPy adapters remain in Python. A later Rust/PyO3 kernel is
appropriate only for stable exact rational constraints, DAG/obligation
transport, bounded candidate enumeration, and certificate replay after the
schema is frozen and profiling justifies FFI/packaging cost.

No main-package dependency or Experimental/Public API is introduced.

Governance

• Mathematical Core: unchanged.
• Research Programme: local pressure on U1/U2/E only.
• Research Status: no repository-wide theorem promotion.
• Engineering Architecture: refined by a research-local symbolic adapter
  before the stable exact balance/replay kernel.
• Theory Map: unchanged; T1/local evidence.
• Universality / 3D Ising / surreal necessity: not earned.

Next claim-expansion gate

Require both:

1. a second independent representation family that is not merely another
   Bessel registry entry; and
2. an executable same-information comparison against Wolfram or a comparably
   strong automatic asymptotics system.

Until then, the responsible disposition is NARROW.

Tracks #140.