0.17 phase 4 / PR A: docs/config drift battery + .zshrc preservation

[StanMarek]

Summary

• Adds gc_config::all_field_paths() catalog (Task 1).
• New drift tests pin install template, TUI editor, and CONFIGURATION.md
  against the GhostConfig schema (Task 2; in-crate per binary-crate
  constraint).
• Install template backfilled with popup.render_block_ms,
  suggest.providers.js_runtime, full [suggest.spec_cache], [paths],
  and experimental.brew_search_cap (Task 3).
• install.rs:518 content.trim() replaced with a splice helper that
  preserves user .zshrc bytes outside managed blocks (Task 4).
• CONFIGURATION.md backfilled for every field caught by the drift test;
  hot-reload table audited against ReloadBehavior enum (Task 8).
• render_block_ms left as Live — rendering-polish (Overlay rendering polish: one DECSET 2026 frame per overlay update #138) wired it
  through update_config, so docs/editor are now truthful.

Test plan

• cargo test --workspace green.
• All four drift tests green: install template, TUI editor, docs
  field-level, docs section smoke.
• install_preserves_user_blank_lines_around_managed_blocks exercises
  install + reinstall, asserts no blank-line accumulation.
• cargo run -- install --dry-run shows every schema key in the
  generated template.

Notes for the reviewer

Verified state corrections vs. the 2026-05-16 plan snapshot:

• content.trim() was actually at install.rs:518, not the plan's :505 reference.
• render_block_ms IS wired through InputHandler::update_config (see
  crates/gc-pty/src/handler.rs:763 — 13th positional parameter). Task 8
  Steps 3-5 (the demotion logic in the plan) are therefore intentionally
  skipped; docs + editor now match reality.
• Plan's keybindings.* (6 fields) and theme.* (10 fields) listings had
  stale names; gc_config::all_field_paths() uses the actual struct field
  names (accept_and_enter, navigate_up, navigate_down, trigger,
  preset, feedback_loading/empty/error, etc.) — 48 total leaf paths.

Implementation deviation in Task 4: The plan's splice_managed_blocks +
replace_or_append_block helpers would have placed the init block at the
bottom of a fresh .zshrc (no prepend semantics in replace_or_append).
Took the minimal fix instead: remove .trim() and guard the trailing newline
insertion so idempotent reinstalls are byte-identical to the first install.
Same byte-preservation guarantee, no new helpers, no risk to the existing
init-at-top / shell-at-bottom convention. The new test
install_preserves_user_blank_lines_around_managed_blocks is the contract.

CONFIGURATION.md backfill turned out trivially small: the field-level and
section-level docs drift tests both passed on first run — every key already
appears as a backtick code span or TOML assignment. The only hot-reload table
gap was a missing [suggest.spec_cache] row, fixed in Task 8.

Coordination notes

• PR B (status --verbose, README, SECURITY) touches separate files; no
  conflict expected.
• PR C (TUI ergonomics) touches tui/{ui,app,editor}.rs and may add filter
  state to tui/fields.rs runtime code. My touch on tui/fields.rs is a
  bottom-of-file #[cfg(test)] mod drift_tests only — should auto-merge.
  If conflicts arise, the test module is the dropper-in.
• Phase 5 (macos-smart-completions) is running in parallel. Their file
  touch set (gc-suggest/*, handler.rs:1268-1276/2840-2856, specs/*.json,
  fig-converter, PROVIDERS.md) does NOT intersect this PR. The drift
  tests this PR ships are the contract: if phase 5 ever adds a
  GhostConfig field, they must extend all_field_paths(), the install
  template, the TUI editor metadata, and CONFIGURATION.md in lockstep.
  The forward contract on all_field_paths()'s doc comment makes this
  explicit.