feat(proxy): honor + echo X-LLMTrace-Trace-Id header (#91 E2E-L1a)

[epappas]

Summary

First concrete step of the E2E adversarial test framework (umbrella #91).

• Pre-flight audit (Loop E2E-L1) identified one real gap: the proxy always generates a server-side trace_id at request entry and never echoes it, so the harness can't correlate per-request observations (findings, metrics deltas, judge verdicts) with scenarios.
• This PR closes that gap (Loop E2E-L1a) with a minimal, typed change and unblocks Loops E2E-L4 (metrics-delta observer) and E2E-L5 (judge verdict collector).

Changes

crates/llmtrace-proxy/src/proxy.rs

• pub const TRACE_ID_HEADER = "x-llmtrace-trace-id" and a new extract_or_generate_trace_id(&HeaderMap) -> Uuid helper. Reads the inbound header, trims whitespace, parses as Uuid. Falls back to Uuid::new_v4() on missing/non-UTF-8/unparseable — preserves current behaviour when no header is sent.
• proxy_handler calls the helper in place of the unconditional Uuid::new_v4().
• Every response path stamps X-LLMTrace-Trace-Id: <uuid>:
  • Success path builder (set after the upstream-header copy loop so we win on conflict)
  • error_response(status, message, trace_id) — new required param; 4 call sites updated
  • rate_limit_response(…, trace_id) — new required param
  • cap_rejected_response(…, trace_id) — new required param

docs/TODO_E2E.md

• New file. Pulls the 10 E2E child issues (feat(e2e): attack-scenario YAML schema + validator #92–feat(e2e): Rust llmtrace redteam subcommand (stretch) #101) into a RALPH-style loop breakdown with IDs (E2E-NNN), acceptance criteria, dependency graph, and sequencing options.
• Includes the Loop E2E-L1 pre-flight audit findings that motivated this PR, plus two other L1 outcomes:
  • JudgeVerdictQuery { trace_id: Option<Uuid>, .. } already exists, so the L5 plan is trimmed (no new find_by_trace_id trait method needed — query_verdicts suffices).
  • Default shutdown.timeout_seconds = 30; L3 e2e fixture should override to 5 so SIGTERM teardown fits the harness's 10-second budget.

Test plan

• cargo test -p llmtrace --lib proxy:: — 27/27 pass (4 new trace-id tests + 2 updated response-shape tests + 21 pre-existing)
• cargo test -p llmtrace --lib — 569/569 pass, no regressions
• cargo clippy -p llmtrace --lib -- -D warnings — clean
• cargo fmt --check -p llmtrace — clean

New tests

• test_extract_or_generate_trace_id_honors_valid_inbound — valid uuid round-trips
• test_extract_or_generate_trace_id_tolerates_surrounding_whitespace
• test_extract_or_generate_trace_id_generates_when_missing — empty headers → fresh v4 each call
• test_extract_or_generate_trace_id_generates_when_unparseable — garbage → fresh v4
• test_error_response_format extended to assert echoed x-llmtrace-trace-id
• test_cap_rejected_response_format extended to assert echoed x-llmtrace-trace-id

Unblocks

• Loop E2E-L4 (E2E-035) — metrics-delta observer can now scope counters to a client-supplied trace_id.
• Loop E2E-L5 (E2E-045) — judge verdict collector can poll by the same trace_id it supplied on the request.

Backward compatibility

• Non-breaking for external callers: if no X-LLMTrace-Trace-Id header is sent, behaviour is identical to before (server-generated UUID).
• New response header is additive; existing x-llmtrace-flagged / x-llmtrace-findings headers unchanged.
• Private helper signatures (error_response, rate_limit_response, cap_rejected_response) gained a required trace_id parameter. These are internal to the crate; no public API change.