Hide init/metrics behind the existing gate pattern, withdraw them from the user guide, correct doctor's checks

[simonkrol]

CLI:

• Hid konductor init and konductor metrics from --help and gated them at dispatch behind KONDUCTOR_ALLOW_INIT=1/KONDUCTOR_ALLOW_METRICS=1 — matches the existing konductor config pattern exactly. Both are complete, tested commands withheld from the v1 customer-visible surface, not removed; metrics is additionally a genuine stub with no logic behind its gate.
• Removed doctor's config check from the live checks vector (run_checks) — matches the existing check_gitignore dormant-check precedent. Traced every consumer of the resolved config values and confirmed none acts on them for a real decision; the check was validating decorative output. Function and its unit tests stay intact, just not wired into live dispatch. doctor now runs 8 checks instead of 9.
• Fixed doctor's dormant config-load-failure remediation string (in check_config_with_home, a check that is unit-tested but never called from live dispatch): it described a backup/hand-edit/delete-to-fallback recovery workflow for config.yml — the same stale "active risk" framing already removed from the docs. Since nothing reads or writes this file through any live command today, there's nothing to remediate against; the string now says so plainly instead of describing a workflow that doesn't apply to anything a user can currently do.
• Corrected cli.rs's --from doc comment, which claimed installing from a published release was unavailable — false since v1.0.0 shipped; now describes the real GitHub-release/main-dist fallback chain.
• Fixed the no---from remote install path printing "skipped 0 SOP(s) (no runtime discovery path yet)" even when it had installed 19 real SOPs. Added a manifest-derived count so the message reports accurately on both the --from and no---from paths.
• Separated install's SOP and skill counts: SOP-derived sop-<name>/SKILL.md conversions no longer count toward the skill total. Detection now looks for the <agent-sop name="..."> marker every conversion carries, rather than the sop- filename prefix alone, which had misclassified a hand-authored skill (skills/sop-state-management/) that happened to share the prefix.
• Corrected cli/README.md's --all section, which still said "all six checks above" after the live check count had already moved to 8
• Added the three update flags (--cli, --version <v>, --force) to reference.md's flag table — present since the v1.0.0 release but never documented there; tasks/update.md already covered them in full.
• cli/README.md brought fully in line with all of the above: command list, status callout, check table, sample outputs, and "Current state" section.

Docs:

• Restructured docs/user-guide/quick-start.md to lead with the curl-based quick install (fetches the published release) instead of framing source-building as the only path. Corrected a real misunderstanding baked into the prior draft — installing without --from never runs synth remotely, it fetches synth's already-built output.
• Restructured docs/user-guide/tasks/update.md so the no---from content-update path (with --version, --force, doctor's new version-checks) leads, with --from demoted to a clearly labeled alternative for maintainers working from a local checkout.
• Deleted docs/user-guide/tasks/initialize-a-project.md and removed every teaching reference to init/metrics from reference.md, concepts.md, glossary.md, faq.md — consistent with both being withdrawn from the documented command surface. Reworded every remediation step that used to recommend running init so it gives a manual alternative instead of a command that will now fail.
• Corrected .konductor/config.yml's documentation in reference.md, notes.md, and tasks/update.md: it is not dead code — its load/validate logic is real and functional — but it is currently unreachable because config (the only command that reads or writes it) is hidden/gated the same way init/metrics are. Removed the "schema change might bite you" risk/recovery messaging entirely (back up your config, hand-edit it, re-scaffold it) since there is no live workflow where that risk currently applies; replaced with a flat statement that the file is not currently used.
• Corrected faq.md's "Is any of my data sent anywhere?" section, which claimed the CLI "touches the network in one place only: install" and that "neither update nor doctor makes any release call" — both false. doctor fetches the latest release tag by default to check currency (skippable via --no-version-check); a no---from update fetches a release by design; update --cli additionally reaches the network for its self-replace step. synth remains the only command with zero network involvement, and is now stated as such specifically rather than folded into a blanket "everything else is local" claim.
• Fixed all navigation left dangling by the initialize-a-project.md deletion (prev/next links, ToC entries, a mermaid diagram).
• Corrected the doctor check-count documentation from nine to eight everywhere it's stated, and fixed a stale "no release has been published yet" claim in prerequisites.md.
• Resynced both docs/index.html and docs/site/user-guide.html to match every markdown change above, via the deterministic htmlbundle.py/sync-html-content.py edit mechanism — confirmed byte-identical to each other throughout, and independently verified by decoding both and diffing against current markdown content (not just the automated checker's narrower fact-based checks).

Build tooling:

• Fixed a real gap in check-guide-facts.py's own withdrawn-command leak detector: its pattern required a trailing space that a JSON-quoted string value never has, letting metrics leak into both live HTML bundles undetected. Fixed the boundary logic generally, not as a metrics-specific patch. Extended it further with a bare-word enumeration-leak check after finding a second leak of the same underlying shape.
• Retired several stale entries in tools/build-user-guide/html_sop_edits.py/sync-html-content.py whose new text encoded facts later corrected elsewhere — each was silently reintroducing already-fixed content on every subsequent --apply. Root-caused and fixed at the source rather than patched again in the HTML.

Release:

• Bumped cli/konductor-rs's Cargo.toml/Cargo.lock and the top-level VERSION file to 1.0.4, with a new [1.0.4] CHANGELOG.md entry. This branch was rebased onto main after PR fix(telemetry): scoped agent telemetry for Kiro v2/v3 and Claude #20 (fix/telemetry-kiro-v2-v3) — earlier expected not to merge — actually merged and landed as the real, published 1.0.3. That PR's own [1.0.3] CHANGELOG section (telemetry hook wiring, mcp/README.md drift) is left untouched above this entry; this release is 1.0.4, not a second claim on 1.0.3.

Fixes # (issue)

Type of Change

• CLI (cli/) change — Rust or Python
• Documentation

Testing

• If cli/ changed: cd cli && make test passes (Rust + Python + conformance suites) — full suite green after the rebase onto PR fix(telemetry): scoped agent telemetry for Kiro v2/v3 and Claude #20's merged, substantially-rewritten cli.rs/dispatch.rs/doctor.rs/install.rs, confirming every fix's intent survived relocation into the new upstream code structure, not just that it compiled
• Smoke tested affected agent(s): live-verified konductor --help, konductor init/metrics/config (gated, exit 64, "not currently available"), and each with its escape hatch set (real behavior confirmed — init wrote real files to a scratch target, metrics reached its stub)
• Ran the benchmark harness for the affected agent(s) — not applicable, no agent behavior changed (CLI/docs only)
• New or changed code files carry the required SPDX header

Checklist

• My code follows the style guidelines of this project
• I have performed a self-review of my own changes
• I have commented my code where necessary
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added or updated tests that prove my change works — added gate/escape-hatch tests for init/metrics mirroring the existing config test pattern; added a regression test for the SOP/skill double-count exclusion (order-independent); added a serializing-lock guard to a test that raced its siblings over the same env var; updated one test whose assertion no longer held once config's doctor check was removed

Notes for Reviewers

• The chore: release 1.0.4 commit (version bump + changelog) is intentionally separate from the substantive fix: commits and is expected to be squashed into the CR or folded into one of them at merge — it's release bookkeeping, not an independent change.
• docs/user-guide/'s version strings are deliberately still pinned at 1.0.0, unrelated to this bump — a standing, documented decision (see notes.md), not an oversight.

[knihit]

Posted comments for clarification.

[simonkrol]

Closing in favor of #31 as fork PRs don't run CodeQL checks.