feat(evals): add mastra bench harness (integrations/mastra-sdk) (#2807)

## What

Adds `--harness mastra` to evals, backed by a new private
`@browserbasehq/stagehand-integrations-mastra-sdk` package.

- `packages/integrations/mastra-sdk` — `runMastraSession`: in-process
`@mastra/core` Agent + `@mastra/mcp` MCPClient (per-session UUID id),
streamed tool-call/result/usage events, abort forwarding,
status/stopReason normalization, redaction.
- `packages/evals/framework/mastraToolAdapter.ts` — `via:"mcp"` mounts
(stagehand_facade, playwright_mcp, chrome_devtools_mcp) with idempotent
cleanup; step observations recorded.
- `packages/evals/framework/harnesses/mastraAdapter.ts` — events →
`NormalizedToolCall[]`.
- Registered via `defineExternalHarness` on the shared external-runner
skeleton from the base PR.

## Testing

- Full unit gate green: evals 488, mastra-sdk 12, all other suites
unchanged.
- Connected smoke (`evals run b:webvoyager --harness mastra --tool
stagehand_facade -l 1 -t 1 -e browserbase`): harness drove **7 Stagehand
tool calls** through the facade and produced a final answer.

Stacked on the wave-core PR. Implemented with Codex (gpt-5.6) under
supervision; Claude + `codex exec review` findings (MCPClient id
collision, env leakage, redaction) fixed in-branch.

<!-- This is an auto-generated description by cubic. -->
---
## Summary by cubic
Adds a selectable `mastra` eval harness backed by the private
`@browserbasehq/stagehand-integrations-mastra-sdk` package. Evals can
now run tasks through `@mastra/core` and `@mastra/mcp`, sharing the
external-runner prompt, parsing, grading, and metrics flow used by the
other agent harnesses.

- Wraps `Agent.stream` and `MCPClient` with per-session client IDs,
abort forwarding, bounded discovery and cleanup, normalized usage, and
sanitized output.
- Mounts `stagehand_facade`, `playwright_mcp`, and `chrome_devtools_mcp`
over stdio MCP, while exposing code surfaces through the existing
in-process bridge.
- Converts Mastra stream events into sanitized trajectories with paired
tool calls and results, image extraction, and correctly concatenated
reasoning and text fragments.
- Treats an SDK error as a failed run even when the model emitted
valid-looking success JSON first.
- Registers `mastra` with `stagehand_facade` as the default tool
surface, `openai/gpt-5.4-mini` as the default model, and
`EVAL_MASTRA_MODELS` as the override.
- Adds workspace dependencies, lockfile and Turbo targets, CI build
caching, and unit coverage for the SDK, adapters, runner, registry, and
planner.
- Mastra requires `@mastra/core` and `@mastra/mcp`; MCP definitions must
use stdio servers with a string `command`.
- Supports step, MCP, evidence-capture, and teardown limits through
`EVAL_MASTRA_MAX_STEPS`, `EVAL_MASTRA_MCP_TIMEOUT_MS`,
`EVAL_CAPTURE_EVIDENCE_TIMEOUT_MS`, and
`EVAL_AGENT_MOUNT_CLEANUP_TIMEOUT_MS`.

<sup>Written for commit b9180990d29d12840ba61f6d876f9270fb5e75f5.
Summary will update on new commits.</sup>

<a
href="https://cubic.dev/pr/browserbase/stagehand/pull/2807?utm_source=github"
target="_blank" rel="noopener noreferrer"
data-no-image-dialog="true"><picture><source
media="(prefers-color-scheme: dark)"
srcset="https://www.cubic.dev/buttons/review-in-cubic-dark.svg"><source
media="(prefers-color-scheme: light)"
srcset="https://www.cubic.dev/buttons/review-in-cubic-light.svg"><img
alt="Review in cubic"
src="https://www.cubic.dev/buttons/review-in-cubic-dark.svg"></picture></a>

<!-- End of auto-generated description by cubic. -->

---------

Co-authored-by: miguel <miguel@browserbase.com>