v1.9.4 — /setup-compile-aware two-phase (cclsp.json audit + compile_aware)

[1.9.4] — 2026-05-22

PATCH release shipping a meaningful expansion of the
/setup-compile-aware skill. The skill is now two-phase:
Phase 0 audits cclsp.json against the project's actual language
files and proposes additive-only LSP entries for languages that
aren't covered yet; Phase 1 keeps the existing
[compile_aware.commands] proposal flow.

Motivation: cclsp.json was previously installer-only, init-only —
written once at install.py time with all detected Tier-1 LSes
wired to all their default extensions, regardless of whether the
project actually used those languages. A Python-only project got
gopls, clangd, rust-analyzer, omnisharp entries it would
never use; a project that grew to include TypeScript six months
later had no skill-driven way to wire up
typescript-language-server. The expanded skill closes that gap.

No Python code changes; no schema or config changes. Pure content
expansion of .claude/skills/setup-compile-aware/SKILL.md (200 →
477 lines) + corresponding docs/lsp-engine.md update.

Test suite: 4015 passed / 136 skipped on Python 3.11 (unchanged
from v1.9.3 — skills are markdown prompts, the v1.9.3
manifest-completeness tests already lock the structural
invariants).

Added — /setup-compile-aware Phase 0: cclsp.json audit

When the skill is triggered, it now walks the project before
touching any file:

1. Reads the existing cclsp.json (if any). When the file is
   missing, Phase 0 is init; when present, Phase 0 is audit. Both
   paths converge on the same additive-only flow.
2. Walks the project for language files via Glob — counts
   .py/.ts/.rs/.go/.c/.cpp/.cs/.sh/.lua/etc,
   skipping vendored / generated dirs (node_modules/, .venv/,
   dist/, build/, target/, __pycache__/, .git/). Uses a
   ≥ 3-file threshold (or a build marker that implies the
   language, e.g. Cargo.toml → Rust regardless of file count)
   so a stray .py file in a Rust project doesn't wire up
   pyright.
3. Computes the gap as three sets:
   • Missing — languages with files in the project but no
     LSP entry yet. These are the additive candidates.
   • Covered — already wired up, preserved verbatim.
   • Strangers — registered LSPs with zero matching files
     in the project. Surfaced for awareness but never
     touched (the file may contain custom flags or
     deliberate user choices the skill can't infer).
4. Probes each proposed binary with command -v and warns if
   missing. Registering an absent binary is not destructive —
   the engine returns [] until the binary lands — but the
   skill surfaces the install hint from
   claude_hooks/lang_servers.py:SPECS so the user can decide.
5. Asks for confirmation with [Y/n/edit] per proposed
   merge. The user can drop specific entries.
6. Backs up + writes — timestamped backup
   (cclsp.json.bak-YYYYMMDD-HHMMSS) precedes any write, then
   the merged JSON (existing entries verbatim, new additions
   appended) lands at cclsp.json with diff-back verification.

Invariant — additive-only on cclsp.json

The skill never proposes removing or modifying an existing
entry. Custom flags (--strict, --remote=auto), manual entries
for niche LSes, and deliberate user choices are preserved
verbatim. Adding new entries to cover languages the user didn't
have before is safe; touching existing ones is not. Stranger
entries (registered LSPs with zero matching files in the project)
are surfaced informationally but left alone — the user can clean
them up manually if vestigial.

Changed — docs/lsp-engine.md

The ## Compile-aware setup section is now ## LSP engine setup
with a phase-by-phase walkthrough of what the skill does and the
additive-only invariant spelled out. The brief Phase-3 mention in
the Compile-aware-diagnostics section now reflects the two-phase
scope.

Upgrade notes

Re-run python install.py on every host. The installer's
_install_skills does a content-comparison between the in-repo
SKILL.md and the user-scope copy — if the content differs (it
will after v1.9.4), it prompts [Y/n] to update. Accept to pick
up the two-phase prompt.

The skill is identical-named (/setup-compile-aware); no
muscle-memory loss. The behavior change is purely additive — if
your project already has a cclsp.json that covers every
language present, Phase 0 reports "nothing to propose" and the
skill behaves exactly as in v1.9.3.