feat(pprof): emit per-sample process_language label

[r1viollet]

What

Emit a per-sample pprof label process_language carrying a heuristic native-language family for the profiled process' main executable (go / rust / cpp). When detection is inconclusive, no label is set and the existing language="native" tag continues to be the sole signal.

Why

Today every native sample is reported as language="native" — Rust, C++, plain C and (mostly) Go are indistinguishable on the backend side. With this label, the backend can refine the per-process language for the dominant case where one language clearly drives a binary, without splitting profiles.

Mirrors the approach we will likely use in the OTel eBPF profiler / Full-Host Profiler (also no DWARF involvement, also per-sample label rather than a global tag), so the two producers will be consistent.

How

• New NativeLanguage enum + cheap detection (include/native_language.hpp / src/native_language.cc):
  • Go → .go.buildinfo / .gopclntab ELF section
  • Rust → .note.rustc section, or rustc-mangled symbols in .symtab/.dynsym (_R… v0, or legacy …17h<16 hex>E tail)
  • Cpp → any _Z… Itanium-mangled symbol
  • else → kUnknown (label omitted)
• Symbol-table scan bounded to 4096 entries.
• Detection runs once per PID, lazily on first sample, against the Elf* already cached by libdwfl (dwfl_module_getelf) — no extra file open.
• Result cached on Process, cleared via the existing ProcessHdr::clear(pid) on process exit.
• A /proc/<pid>/exe fallback path is kept for callers without an Elf* in hand (tests / early bring-up).
• Emitted as a pprof per-sample label process_language alongside process_id, process_name, etc. (k_max_pprof_labels raised 8 → 9).

Verification

Check	Result
tools/style-check.sh	clean
Full ninja (853 targets)	passes
ctest -j4	52/55 pass; the 3 failures (ddprof_stats-ut, ipc-ut, simple_malloc-with-event-reordering) reproduce on main without this patch — pre-existing, unrelated

Smoke-tested end-to-end with ddprof -p <pid>:

• Busy Go binary → [NATIVE-LANG] -> go (single detection log across all samples; libdwfl-cached Elf* reused)
• C++ binary → cpp
• /bin/cat, sleep (C) → kUnknown → no label, falls back to native tag

Notes for review

• Heuristic by design. Mixed-language processes (e.g. C++ with linked Rust crates) will resolve to whichever signal hits first (Rust wins over C++). Acceptable for the dominant-language reporting goal.
• No DWARF reads. Symbol tables are scanned via libelf only; cost is one-shot per FileID and bounded.
• No new dependency. Reuses already-linked elfutils.
• Per-sample label vs global tag. Picked per-sample so multi-PID profiling stays correct; the global language="native" tag is unchanged.
• Symbol-table availability. For fully stripped binaries with no .dynsym C++ symbols, Cpp detection may miss — expected, label stays unset → native.

[r1viollet]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. 🎉

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".