refactor: shrink fallback, merge email benchmarks, dedupe bool prefs

[qasimsethi-code]

Ran /ponytail-audit on the repo and applied the surviving findings. Net −41 lines, no behavior changes except one deliberate fix (last bullet).

• getFallbackInstructions: condensed the 31-line hand-maintained SKILL.md mirror to paragraph form — the same shape _fallback_instructions in the Hermes port already uses. It only runs when SKILL.md (shipped in the same package) is unreadable; the full mirror was sync debt with no payoff.
• benchmarks/email-rate.js replaces claude-email.js + model-email.js: same harness loop existed twice differing only in the call() and env names. One file, two-entry provider table. Env contract: PROVIDER=claude|openai, N, MODELS (was CE_*/ME_*).
• .env parsing: hand-rolled kv parser (duplicated in both email scripts) → util.parseEnv (Node ≥20.12).
• boolPref(envVar, configKey) in ponytail-config folds the twin getQuietStartup/getHideStatus resolvers; exports and env→config→default semantics unchanged.
• Hermes port fix: _filter_skill_body_for_mode's example regex was missing the " quote guard the JS version added deliberately — a prose bullet starting with a mode word was silently dropped in every other mode. Now mirrors hooks/ponytail-instructions.js.

Audit also flagged the root plugin.json as unreferenced — wrong, it's the Grok adapter manifest (tests/grok-plugin.test.js caught it); left untouched.

Tests: root 83/84 (the one failure is csv: correct pandas one-liner passes, environmental — no pandas on this machine, fails on a pristine checkout too), hermes 8/8, pi-extension 23/23, ponytail-mcp 3/3.

🤖 Generated with Claude Code

Fixes #772.

[qasimsethi-code]

Friendly triage note, no rush at all: this is primarily a bugfix for #772 (the Hermes port silently drops prose bullets that start with a mode word), with the audit cleanups riding along. Since first-contribution CI needs maintainer approval here, I ran the identical test workflow on a temporary mirror — passed clean on ubuntu-latest / Node 22 / Python 3.12 (84/84 across all four suites). Happy to split the bugfix into its own minimal PR if that's easier to take.