Round-2 patch for autoresearch

Three files updated to close the loopholes the round-1 agent found:

File	What changed
autoresearch/invariants.mts	+6 new locks. Total: 19 → 25 invariants.
autoresearch/program.md	New tier-ranked hypothesis seed list; explicit forbidden-as-hypothesis list.
autoresearch/calibration/cases.json	8 → 28 cases. Adds W3/W4/W5 regime cases, EDR boundary, equipment block/warn, operational context, low-forecast-confidence, climate trend.

How to apply

From the openlarm/larm repo root:

# back up your round-1 results first
cp autoresearch/results.jsonl autoresearch/results.round1.jsonl

# revert the round-1 v2.ts (those changes had structural problems)
git checkout main -- packages/regions-taiwan/src/v2.ts

# apply this patch
unzip ~/Downloads/openlarm-autoresearch-round2.zip

# clear results log for the new run
echo '' > autoresearch/results.jsonl

# verify baseline
node autoresearch/evaluate.mjs

You should see something like:

build             : PASS
conformance       : PASS
invariants_passed : true (25/25)
metric_score      : 0.5076   ← new baseline on 28 cases
worst_miss        : CAL-015-edr-just-below-hard-stop (pred GO/R1, exp NO_GO/R3)
elapsed_sec       : ~11s

What's new in invariants.mts

1. mapping_r_level_canonical_spec_boundaries — locks R-level boundaries to canonical spec values (R0:0–20, R1:21–40, R2:41–65, R3:66–85, R4:86–100). This blocks the cheapest round-1 metric hack.

2. wr_matrix_w0_low_risk_not_nogo — W0 (stable clear weather) R0/R1/R2 cells must not be nogo. Blocks round-1 exp 14 hack.

3. wr_matrix_r4_always_nogo — every regime row must end in nogo at R4.

4. wr_matrix_no_go_above_r1 — "go" decisions can only appear at R0/R1, never R2/R3/R4.

5. regime_base_score_ordering — W0 must be lowest, W5 highest. Stops the agent from re-ordering regime severities to redirect cases through different W-rows.

6. rain_hard_stop_thresholds_meaningful — keeps rain hard-stop thresholds within bands defensible against thunderstorm cases.

What's new in program.md

The hypothesis seed list is now 5-tiered with explicit ordering:

• Tier 1 (try first): weather_now_weights, wind_score_table row scores, rain_score_rules — the actual sensitivity knobs the loop was designed for
• Tier 2: EDR thresholds and adjustments
• Tier 3: regime base scores, region_weight_table, volatility_buffer_add
• Tier 4: wr_matrix cells (one at a time, with justification)
• Tier 5: buffer ratios, component caps

The ## EXPLICITLY FORBIDDEN as hypothesis section now names every round-1 metric hack pattern with the exact invariant that catches it. Agent will stop trying these on iteration 1 instead of grinding through 5–6 revert cycles per pattern.

What's new in cases.json

20 new cases covering gaps I noticed in round 1:

ID	Tests
CAL-009	W3 meiyu regime with a clear-window operational decision
CAL-010	W4 climate but morning mission — temporal awareness
CAL-011	Light rain at low probability — should stay GO
CAL-012	Rain rate 11mm/h at 40% prob — must NOT trigger conjunction hard stop
CAL-013	Coastal windward exposure under NE monsoon
CAL-014	EDR=0.45 borderline turbulence
CAL-015	EDR=0.78 just-below hard stop (current worst_miss)
CAL-016	Wind 38 km/h just below 39 hard stop
CAL-017	Equipment block status
CAL-018	Two equipment items in warn
CAL-019	Tight clearance, small building
CAL-020	Multi-day split mission
CAL-021	Weekend + crowded area + road closure
CAL-022	Near 5G base station
CAL-023	W5 with high recent_typhoon_count
CAL-024	w_override regime (manual W4 forcing)
CAL-025	Low forecast confidence — buffer_ratio test
CAL-026	Poor rooftop condition
CAL-027	W0 climate + sudden high wind today
CAL-028	Urgent deadline pressure

Note CAL-015 and CAL-006: round-1 agent identified these as structurally unfixable by params (engine caps and hardcoded R3-CONDITIONAL behavior). With the new invariants in place, if these stay as worst_miss after round 2, that's evidence to take them to the engine (Code TSC review) rather than try more parameter tweaks.

Restart prompt for the agent

Read autoresearch/program.md (round 2). The R-level boundary lock is
new — do not waste experiments trying to move them. Focus on Tier 1
hypotheses (weather_now_weights, wind/rain score tables) first. Run
30 experiments. Append to autoresearch/results.jsonl. Revert any
change that fails build, conformance, invariants, or doesn't strictly
improve metric_score.

If the agent runs into a metric ceiling around 0.7–0.8 and the worst_miss stays on CAL-006 / CAL-015 / R3-hardcoded cases, stop the loop and write up findings — those are engine-level concerns that need human Code TSC review per OPEN_SOURCE_DECISIONS.md, not more calibration iterations.