fix(installer): failed index build no longer reports install success (#351)

[HumanBean17]

What

java-codebase-rag install returned exit 0 and reported success even when the indexing step (cocoindex / graph build) failed. The cocoindex failure was swallowed and install looked healthy to CI/automation.

Note on the issue's repro

Issue #351 cited two red tests (test_install_emits_indexing_progress_on_stderr, test_install_indexing_exception_renders_failed_footer). Both already pass on current master — the progress/footer machinery they exercise was added after the issue was filed. But the core bug remains in the code: run_install still discards run_init_if_needed(...) and unconditionally return 0 (installer.py). The untested path is a non-exception failure — cocoindex returns returncode != 0 (no raise) → run_init_if_needed returns False → discarded → exit 0.

Fix

run_init_if_needed's False conflated two outcomes: "ran and failed" and "skipped because the index already exists". A naive return 0 if init_ok else 1 (the issue's literal suggestion) would make idempotent re-runs (existing index) wrongly exit 1. So:

• run_init_if_needed now returns True (ran OK), False (ran and failed), or None (skipped — index exists). It has exactly one caller (run_install).
• run_install captures the outcome and returns 1 only when it is False; a skip (None) or success (True) stays 0 — mirroring run_update's if not index_ok: return 1.

Test

New test_install_indexing_failure_returns_nonzero patches run_cocoindex_update to return returncode=1 (non-exception failure) and asserts run_install exits 1 with CocoIndex update failed on stderr. TDD: red first (exit 0 despite cocoindex failure), then green.

.venv/bin/python -m pytest tests/test_installer.py -q
# 84 passed

Notes

• No re-index, env-var, or ontology_version change — installer lifecycle only.
• Behavior change: install now exits non-zero (1) when indexing fails. Idempotent re-runs over an existing index still exit 0.

Closes #351.

🤖 Generated with Claude Code

[HumanBean17]

Addressed 2 review findings:

1. Skip→None→exit-0 untested — added test_install_over_existing_index_skips_init_and_exits_zero: a re-run over an existing index skips init and exits 0. Guards the PR's central if init_outcome is False design so a future if not init_outcome simplification (collapsing None into the failure branch) fails this test. Verified red→green.
2. install reports success (exit 0) on failed index build; swallows cocoindex exception #351 deferred scope — documented at the partial-failure classifier that "treat skill/agent deploy failures as critical" is intentionally deferred (only MCP-config failures are critical: fatal vs. recoverable), so it reads as a deliberate decision, not an oversight.