docs: complete optimizer variants map and dose semantics

[charlesmartin14]

Status

Ready to merge. This PR supersedes #31.

PR #34 has been merged. This branch was rebuilt directly from the resulting main tip, so the effective diff is documentation-only and contains exactly:

• OPTIMIZER_VARIANTS.md
• README.md

What changed

• Adds an authoritative optimizer map covering all seven current experiment packages.
• Separates actuator, support/spectral geometry, and intervention cadence.
• Adds the current full_matrix_log_rg active-set cone projector and projected Nesterov-state semantics.
• Corrects the trace-log description: the actuator removes flow along a trace-log normal evaluated on retained support; it does not project accumulated weights into an ECS.
• Replaces the ambiguous rank exponent q with mu_rank, reserving q/Q for the Marchenko–Pastur aspect ratio.
• States the density-to-rank exponent correspondence with its assumptions and finite-size qualifications.
• Documents the exact wwpgd_local_delta realized dose:

||Delta_removed||_F / max(||Delta_epoch||_F, eps)

• Distinguishes realized dose from controls and caps such as correction_fraction and projection_strength.
• Replaces the stale partial optimizer list in the root README with a link to the complete authoritative map.
• Removes undocumented claims about external nanoGPT or WW-PGD adapter telemetry.

Scope

Documentation only. No optimizer code, ECS selection, defaults, hyperparameters, experiment recipes, or efficacy claims are changed.