refactor: remove redundant feature gates, promote index into full

[bug-ops]

Summary

Audited all 184 Cargo feature keys across the workspace (root Cargo.toml + 28 crates), classified each as either an unconditional/always-on candidate or a genuinely required modularity boundary, and implemented the ratified classification:

• Removed 27 feature keys (184 -> 157) that provided no real build-time or binary-size benefit: empty markers (deep-link, cocoon), keys already unconditionally enabled by non-optional dependencies (llm-planning, zeph-core/sysinfo), and single-tiny-subfeature gates (registry). Fixed zeph-skills's miner feature to declare explicit dep: prefixes instead of silently synthesizing 5 implicit feature keys.
• Added index to the desktop/full bundles instead of removing its gate — it carries real per-crate dependency weight (+43 packages for zeph-agent-context) that removing the gate would have unconditionally charged to every consumer, including published-crate users. This is the first time full compiles the zeph-index integration; fixed an E0063 in zeph-agent-context test code the new combination exposed.
• Dropped testing from the full bundle (breaking change) — full drives release-build-full, which was silently shipping mock LLM providers in release binaries. Consumers relying on --features full for test doubles must now request testing explicitly.

Kept gated, with reasoning verified per-crate (not just at the workspace-root binary, which hides published-crate consumer cost): sqlite/postgres and metal/cuda (mutually exclusive backends, sqlite/postgres enforced by compile_error!), profiling (spec 035 FR-002 requires it compile to a no-op when disabled), zeph-skills/qdrant, zeph-a2a/{card-signing,ibct,server}, zeph-core/scheduler (load-bearing propagation marker), session (coverage status is Partial, deferred out of this PR), zeph-acp/unstable-cancel-request (deliberate opt-in per #5362).

Also: fixed 6 CI workflows whose hand-copied feature strings had drifted from Cargo.toml, amended specs 001/029 to record the new classification rules and bundle composition, rewrote the stale book/src/reference/feature-flags.md page, and fixed config/default.toml and book/src/advanced/acp.md, which referenced now-removed flags.

Net effect: 0 added package-dependency cost for the entire removed-gate set.

Process

Produced via a 3-round architect/critic classification pass before any code was touched (.local/plan/feature-flags-classification.md, v3 — ratified, 0 open questions), followed by implementation, parallel test-coverage + adversarial-critique validation, and code review with one fix-review cycle (2 blocking documentation issues found and fixed: stale config/default.toml flag references, stale book/src/advanced/acp.md unstable-flag examples that would hard-fail with unknown feature).

Test plan

• cargo +nightly fmt --check
• cargo clippy --profile ci --workspace --all-targets --features "desktop,ide,server,chat,pdf,scheduler,testing" -- -D warnings
• cargo nextest run --workspace --lib --bins on default features (scheduler,sqlite) — 14879/14879 passed
• cargo nextest run --workspace --features full --lib --bins — 15608/15608 passed
• cargo test --doc --workspace --features "desktop,ide,server,chat,pdf,scheduler"
• RUSTFLAGS="-D warnings" cargo check --workspace --all-targets --features desktop,ide,server,chat,pdf,scheduler --locked
• Rustdoc gate (RUSTFLAGS="-D warnings" RUSTDOCFLAGS="--deny rustdoc::broken_intra_doc_links" cargo doc --no-deps --workspace ...)
• --no-default-features --features full,sqlite and full,postgres both compile clean
• sqlite/postgres mutual-exclusion compile_error! guard re-verified (fires correctly both-enabled and neither-enabled)
• cargo metadata --no-deps reports exactly 157 feature keys post-refactor
• Live smoke test: zeph doctor with --features full and [index] enabled = true
• Dead-cfg sweep: zero leftover references to any of the 27 removed feature keys
• Testing playbook + coverage-status rows added (.local/testing/playbooks/feature-flags.md, .local/testing/coverage-status.md)

Two pre-existing, out-of-scope issues surfaced during validation (not fixed here, filed for follow-up): a zeph-core rustdoc failure specific to the untested profiling-alloc combination, and an orphaned zeph-experiments/mock feature with no code consumer.