Fix reflected XSS in the MCP docs page - #50

## Summary

Routine dependency refresh. The most important part is the
`@tigerdata/mcp-boilerplate` bump to `1.6.5`, which pulls in a reflected
XSS fix that affects this server.

## Security fix: reflected XSS in the HTML docs page

`@tigerdata/mcp-boilerplate` `1.6.4` → `1.6.5` picks up
[timescale/mcp-boilerplate-node#50](timescale/mcp-boilerplate-node#50).

The HTML docs page served by `mcpRouterFactory` (`GET` on the MCP mount
path with `Accept: text/html`) built its displayed URL from
client-controlled headers (`X-Forwarded-Proto`, `X-Forwarded-Host`,
`X-Original-Request-Uri`, `X-Original-Uri`) and interpolated them into
the response with no escaping. A crafted header could close the
`<pre><code>` block early and land live markup in the page.

This repo is affected: `src/httpServer.ts` uses `httpServerFactory`, so
it serves that docs page. The upstream fix escapes every interpolated
value with `escape-html` and adds `firstHeaderValue()` so a repeated
header yields the first value instead of a comma-joined list. Verified
the escaping is present in the installed `1.6.5`.

Per the upstream PR, real-world impact was limited: the forwarded
headers are trusted by design so the page can print the externally
visible URL behind the ingress, and the remaining effect is that a
self-directed request can show a misleading URL in its own response
body. No stored or cross-user impact. Upstream flagged host
allow-listing and a CSP header on that response as reasonable
follow-ups.

## Dependency updates

| Package | From | To |
|---|---|---|
| `@tigerdata/mcp-boilerplate` | 1.6.4 | 1.6.5 |
| `ai` | 7.0.2 | 7.0.87 |
| `@ai-sdk/openai` | 4.0.0 | 4.0.53 |
| `zod` | ^4.3.6 | ^4.5.4 |
| `pg` | ^8.22.0 | ^8.23.0 |
| `@biomejs/biome` | ^2.5.1 | ^2.5.11 |
| `@types/bun` | ^1.3.14 | ^1.4.0 |
| `@types/node` | ^24.13.2 | ^24.13.3 |
| `@types/pg` | ^8.20.0 | ^8.23.1 |
| `@types/semver` | ^7.7.1 | ^7.8.0 |

Also bumps bun `1.3.14` → `1.4.0` (the `./bun` launcher and the
`oven/bun` base image in `Dockerfile`, kept in sync).

Deliberately held back:

- `typescript` stays at `6.0.3`. That is the latest stable 6.x; `7.0.2`
is a major.
- `@types/node` stays within 24.x to match the Node 24 target. Latest is
`26.4.0`.

## Breaking changes reviewed

**zod 4.4.0 was the only release with real risk.** It shipped a set of
"potentially breaking bug fixes". Two were relevant here:

- **Floating-point `multipleOf`.** This directly affects
`semanticWeight: z.number().multipleOf(0.1)` in
`src/apis/searchDocs.ts`. No existing test covers non-trivial decimals,
so this was verified by hand rather than trusted to the suite. The fix
works in our favor: `0.3` and `0.7` (the classic float-precision failure
cases, and `0.7` is our documented default) now validate correctly,
while `0.15` and `0.35` still correctly reject.
- **JSON Schema generation.** Confirmed the emitted `inputSchema` and
`outputSchema` are unchanged, so there is no MCP client contract change.
Checked against a live `tools/list` response.

The other 4.4.0 fixes (tuple defaults, `.merge()` refinements, stricter
string validators, record key transforms) do not apply. This code only
uses `z.enum`, `z.string`, `z.number`, `z.coerce.number`, `z.object`,
`z.array`, and `z.union`.

**AI SDK:** reviewed the full 7.x changelog range. The `embed` /
`openai.embedding` surface used by `embedQueryJson` is unchanged. The
one embedding-related change splits OpenAI requests by a UTF-8 byte
budget, which is internal batching and only affects `embedMany`, not our
single-value `embed`.

## Verification

- `./check` passes.
- Full `./bun test` run: 86/86 pass. Note that `./check` runs tests with
`--only-failures`, so a green `./check` alone does not prove the suite
ran; it was run separately.
- `./bun run build` succeeds.
- Smoke-tested the stdio server end to end: migrations run, `initialize`
succeeds, and both `search_docs` and `view_skill` list with the expected
schemas.