feat: honest peer resolution and relative-to-industry confidence - #13

Fixes the reported "3. Relative to Industry — Not available. No curated peer group matched, and the relative module returned zero metrics with Low confidence (0.37, 0% field coverage)".

The bug

peers._group_for matches by exact ticker membership. KPIT was in no group — and neither were Tata Elxsi, Tata Technologies, LTTS or Cyient; the whole automotive ER&D cohort was absent from the repo. With no peers, relative.py skipped all 7 metrics, and evidence.py then computed:

0.80 (Yahoo) × (0.4 + 0.6 × 0.0) × 1.0 = 0.32   +0.05 corroboration → 0.37 "Low"
reason: "source: Yahoo Finance, Curated (Investo); 0% field coverage; cross-source agreement"

That reason claims two sources cross-checked each other over zero rows. The 0.37 wasn't a low-confidence answer, it was a confident-looking number manufactured from nothing — and it leaked into both report.py's and thesis.py's aggregates as a plain mean. A plausible number from no data is worse than a zero, because a zero is obviously a zero.

⚠️ This reprices published numbers

Deliberate, and the main thing to review:

The Buffett path was hit by the identical arithmetic. Report-level confidence will move for affected tickers — upward, since an empty module no longer drags the mean down.

The fix

• Zero coverage earns zero. The 0.4 floor exists so a couple of gaps don't collapse a score; it no longer applies when nothing was computed. The agreement bonus is gated on there being something to agree on.
• A resolution ladder — curated → keyword match on Yahoo's industry → Finnhub → none. The PeerBasis travels with the comparison and discounts confidence through a new reliability_factor, so a guessed cohort can't read like a deliberate one, and a rank among five hand-picked names can never reach the High tier (curated caps at 0.90 by construction).
• auto_erd + auto_components, hospitals_diagnostics, capital_goods_defence.
• A group reframes the narrative, not just outlook/CAGR — Yahoo calls KPIT "Software - Infrastructure" and Cyient "Conglomerates", which points the whole analysis at the wrong drivers. Yahoo's raw string is preserved alongside; hiding the disagreement would be worse than showing it.
• 10 metrics (+EV/EBITDA, ROA, P/S), with coverage measured against what the peer set can actually rank on — otherwise every Indian company silently loses confidence for a metric Indian peers don't report. RelativeMetric.unit stops renderers guessing by name and printing EV/EBITDA as 3000%.

Dead tickers found en route

Each one silently dropped a company out of its own peer table — the reported symptom, arriving by a different route. Found by scripts/validate_peers.py, added here because no offline test can catch it:

• TATAMOTORS.NS — superseded by the TMCV/TMPV demerger; both were already in the group
• SPICEJET.NS → SPICEJET.BO (NSE line gone from the provider)
• LTIM.NS, AKZOINDIA.NS — unresolvable across .NS and .BO; removed rather than guessed at

Verification

ruff + mypy clean; 142 tests pass (103 → 142). test_peers.py was making a live network call on the unknown-ticker path; it's now monkeypatched.

Live, on the actual reported case:

basis=curated  group='Automotive ER&D'  metrics=10  coverage=1.0
confidence=0.77 Medium   (was 0.37 with 0 metrics)

  Metric                  KPIT  Industry   Standing
  ROE                    19.7%     17.1%   top quartile
  Revenue growth         12.0%      4.6%   top quartile
  P/E                    23.7x     31.3x   top quartile
  EV/EBITDA              10.8x     22.4x   top quartile
  Debt/Equity             0.2x      0.1x   below median

Reliance still resolves to oil_gas_energy and Eicher to auto_oem (both are double-booked across groups; a test now pins the first-match-wins order, since appending vs inserting in peers.yaml silently reframes companies).