fix(models): rewrite the maxTokens rationale — my justification was factually wrong

The numbers were right, the code was right, and the reason I gave for both was
wrong. Verified every part of the review's finding first-hand before rewriting.

WHAT I CLAIMED: OpenClaw caps outbound requests via
`maxTokens = Math.min(baseMaxTokens + thinkingBudget, modelMaxTokens)`, so a
16384 declaration at `high` reasoning left ~1024 tokens for the visible answer,
and raising it would produce longer answers and higher quota burn.

WHY THAT CANNOT HAPPEN FOR OCP — three independent breaks, each confirmed here:

  scripts/sync-openclaw.mjs:75   api: "openai-completions"
The clamp I cited lives behind `case "anthropic-messages"`; the
openai-completions transport has no thinkingBudget concept at all, so that code
never runs for a claude-local provider.

  buildCliArgs pushes: --input-format, --disallowedTools, --allowedTools,
                       --dangerously-skip-permissions, --mcp-config
No output-token flag of any kind reaches the CLI. OCP cannot cap output even if
it wanted to.

  grep max_completion_tokens server.mjs lib/  ->  0 occurrences
That is the field OpenClaw would send to a local openai-completions endpoint,
and this repo never reads it.

So "expect longer answers and higher per-request quota burn" was a promise that
would not have materialised. That is worse than a wrong number: a release note
that tells operators to expect a behavioural change which cannot occur.

The change still stands, on the honest reason: models.json is the SPOT (ADR
0003) and its values should be true. maxTokens is ADVERTISED metadata consumed
only by clients that choose to honour it. CHANGELOG and the test comment now
say exactly that, and say plainly that OCP behavior is unchanged.

Test replaced, not just re-worded. The old one asserted `maxTokens > 16384`,
which review showed lets every entry sit at 16385 and still call itself
"registry-aligned" — the actual claim had no coverage. Now pinned per model
against a recorded table of CLI 2.1.220 id-anchored values, with an explicit
failure for any model missing a row. Mutation-proven on both holes:

  all entries at 16385          -> CAUGHT  (the old test passed this)
  sonnet-4-6 wrongly at 64000   -> CAUGHT

Also fixed ocp-connect's unknown-id fallback (8192 -> 32000). Review flagged it
as contradicting the invariant: 32000 is the LOWEST max_output_tokens.default
in the registry, so it is the safe floor and no longer sits below every family
in its own table. Unreachable today, wrong to leave inconsistent.

Verified: bash -n on ocp-connect and py_compile on its embedded python block.
Tests: 458 passed, 0 failed.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017gbqUZ8HfBZpjjbzQ85oH8

CHANGELOG.md

@@ -4,7 +4,7 @@
 
 ### Changed
 
-- **`maxTokens` now matches the CLI registry instead of a uniform 16384 (#195).** Every Opus entry and `claude-sonnet-5` go to **64000**, `claude-sonnet-4-6` and `claude-haiku-4-5` to **32000** — the `max_output_tokens.default` each model actually declares in the compiled CLI 2.1.220 registry. OCP never enforced this value; it is propagated to OpenClaw, where it **caps the outbound request**: `maxTokens = Math.min(baseMaxTokens + thinkingBudget, modelMaxTokens)`. OpenClaw's reasoning budgets are medium 8192 / high 16384, and when `maxTokens <= thinkingBudget` it clamps thinking to `maxTokens - 1024` — so a model declared at 16384 running at **high** reasoning spent ~15360 on thinking and left **~1024 tokens for the actual answer**. `ocp-connect`'s independent family table moves to the family floor (opus 64000, sonnet 32000, haiku 32000) since prefixes cannot distinguish versions. Expect longer answers and correspondingly higher per-request quota burn on long generations.
+- **`maxTokens` now matches the CLI registry instead of a uniform 16384 (#195).** Every Opus entry and `claude-sonnet-5` go to **64000**, `claude-sonnet-4-6` and `claude-haiku-4-5` to **32000** — the `max_output_tokens.default` each model declares in the compiled CLI 2.1.220 registry. This corrects **advertised metadata only**: `models.json` is the SPOT (ADR 0003) and every value in it should be the truth about the model. **It changes nothing about how OCP behaves.** OCP never enforces `maxTokens` — `buildCliArgs` passes no output-token flag to the CLI at all — and OpenClaw addresses a local OCP over `openai-completions`, whose request field (`max_completion_tokens`) appears nowhere in this repo. The value is consumed only by clients that choose to honour it, via `setup.mjs` / `scripts/sync-openclaw.mjs` / `ocp-connect`. **Expect no change in answer length or quota burn.** `ocp-connect`'s independent family table moves to the family floor (opus 64000, sonnet 32000, haiku 32000) since prefixes cannot distinguish versions.
 
 ## v3.25.0 — 2026-07-27