vault-cli: flip analyze default to Opus, --haiku/--sonnet step down

Opus draws the auto-fix/flag line correctly (Haiku over-fixed on judgment
calls, verified apples-to-apples against the same vault baseline). At ~$0.14/run
it's cheap enough to run nightly. --sonnet/--haiku now step DOWN; --opus kept as
an explicit identity flag; raw -m/--provider still override entirely.

Hermes host defaults (config.yaml default preset, delegation/compression models)
are unchanged — this only affects which model the vault CLI's own hermes -z spawn uses.

vault

@@ -15,11 +15,11 @@ Subcommands:
   vault sync-data               reconcile /data skeleton to match manifest (create missing dirs)
   vault manifest                regenerate the topology manifest note
   vault report [--pretty]       deterministic Librarian report (delegates to vault-librarian.py)
-  vault analyze [--sonnet|--opus] [-m MODEL] [--provider PROVIDER]
+  vault analyze [--haiku|--sonnet|--opus] [-m MODEL] [--provider PROVIDER]
-                                spawn a Hermes Librarian pass. DEFAULT: Haiku/Bedrock
+                                spawn a Hermes Librarian pass. DEFAULT: Opus (smart —
-                                (cheap+fast — the nightly/weekly tier). --sonnet / --opus
+                                draws the auto-fix/flag line correctly, ~$0.14/run so cheap
-                                step up intelligence & cost for harder judgment calls.
+                                enough for nightly). --sonnet / --haiku step DOWN for
-                                -m/--provider override entirely (future/non-Bedrock models).
+                                cheaper/faster passes; -m/--provider override entirely.
 """
 import os, re, sys, subprocess
 
@@ -32,9 +32,9 @@ HERMES = "/run/current-system/sw/bin/hermes"
 LIBRARIAN = os.path.join(_SELF_DIR, "vault-librarian.py")
 PROMPT = os.path.join(_SELF_DIR, "vault-librarian-prompt.md")
 
-# `vault analyze` model tiers. Default is Haiku — cheap+fast for the nightly/weekly
+# `vault analyze` model tiers. Default is Opus — smart enough to draw the auto-fix/
-# pass. --sonnet / --opus step up intelligence & cost inline. IDs verified against
+# flag line correctly, and ~$0.14/run so cheap enough for a nightly pass. --sonnet /
-# this host's Hermes config + the active Bedrock provisioning.
+# --haiku step DOWN. IDs verified against this host's Hermes config + Bedrock provisioning.
 ANALYZE_TIERS = {
     "haiku":  "us.anthropic.claude-haiku-4-5-20251001-v1:0",
     "sonnet": "global.anthropic.claude-sonnet-5",
@@ -148,12 +148,15 @@ def main():
         sys.exit(subprocess.run([sys.executable, LIBRARIAN] + argv).returncode)
     elif cmd == "analyze":
         prompt = open(PROMPT, encoding="utf-8").read()
-        # Model tier: default Haiku (cheap nightly). --sonnet/--opus step up.
+        # Model tier: default Opus (smart — draws the auto-fix/flag line correctly,
+        # ~$0.14/run so cheap enough to run nightly). --haiku/--sonnet step DOWN.
         # Raw -m/--provider override the tier entirely (escape hatch for future models
         # or non-Bedrock providers). Tier flags are last-wins if combined.
-        tier = "haiku"
+        tier = "opus"
         if "--sonnet" in argv:
             tier = "sonnet"
+        if "--haiku" in argv:
+            tier = "haiku"
         if "--opus" in argv:
             tier = "opus"
         model = ANALYZE_TIERS[tier]
@@ -179,8 +182,8 @@ def main():
         print(f"  model: {model}", file=sys.stderr)
         print("  output is buffered; a full pass can take 1–3 min before you see anything.",
               file=sys.stderr)
-        if tier == "haiku":
+        if tier == "opus":
-            print("  step up:  vault analyze --sonnet   (or --opus)  for harder judgment calls",
+            print("  step DOWN:  vault analyze --sonnet  (or --haiku)  for cheaper/faster passes",
                   file=sys.stderr)
         sys.stderr.flush()
         subprocess.run(hcmd)