Token efficiency improvements and .c++ extension fix

[andreasjansson]

Changes

fix: recognize .c++ and .h++ file extensions as C++

Files with the .c++ extension (used extensively in codebases like Cloudflare's edgeworker with 577 C++ implementation files) were falling through to "plaintext" and being completely skipped by leta. This made them invisible to grep, graph, show, refs, and all other commands.

docs: recommend leta files before leta graph for exploring projects

Updated the skill and README to use a files-first workflow:

• Start with leta files to understand project size and structure
• Only use leta graph if the project has fewer than ~500 source files
• For larger projects, use targeted leta calls or leta grep instead

feat: make leta graph output more token-efficient

Two changes to reduce token usage in graph output:

1. Relative paths: when a callee is in the same directory as its caller, show ./filename instead of the full path
2. Compact default: omit [Kind] and (signature) by default; use --include-signature to show the verbose format

Before:

crates/core/flags/defs.rs:1537 [Method] update (fn(&self, v: FlagValue) -> Result<()>)
├── crates/core/flags/mod.rs:294 [Function] unwrap_value (fn unwrap_value(self) -> OsString)
└── crates/core/flags/defs.rs:7602 [Function] human_readable_usize (pub(super) fn ...)

After:

crates/core/flags/defs.rs:1537 update
├── ./mod.rs:294 unwrap_value
└── ./defs.rs:7602 human_readable_usize

feat: detect rust-toolchain.toml mismatch and suggest fix

When rust-analyzer fails to start and the workspace has a rust-toolchain.toml or rust-toolchain file pinning to a toolchain that doesn't have rust-analyzer installed, leta now shows an actionable error message:

This project pins to Rust toolchain '1.91' (via rust-toolchain.toml),
which doesn't have rust-analyzer installed.

To fix, run:
  rustup component add rust-analyzer --toolchain 1.91

Previously this showed the generic "install rust-analyzer" message, which was confusing since rust-analyzer was already installed for the default toolchain.

fix: don't cache empty symbol results from LSP

When a language server isn't fully ready (e.g. rust-analyzer still indexing a large workspace), documentSymbol may return empty results. Previously these were cached permanently in the LMDB symbol cache (keyed by file path + mtime), making symbols invisible on all subsequent runs — even after the server finished indexing, and even across daemon restarts.

Now empty symbol results are never cached. The cost of re-fetching for genuinely empty files is negligible.