feat(linearbot): per-turn-type reasoning effort

[0xAlcibiades]

Closes #1557.

Part of #1512 — direction 1, for linearbot. See the scope note at the end.

Change

LINEARBOT_EFFORT_ASSIGNMENT and LINEARBOT_EFFORT_COMMENT (chart:
linearbot.reasoningEffort.assignment / .comment) set reasoning effort per
turn type, threaded through the reasoning field that already exists on the
wire and already maps onto codex turn/start.effort.

Both unset keep today's behaviour exactly — no reasoning is emitted and the
harness default applies. Setting one does not imply the other: a deployment
that wants deep assignment turns has not thereby asked for deep comment
replies, and a test pins that.

Two decisions

Unrecognised values are dropped, not forwarded. A typo forwarded verbatim
reaches the harness as an invalid turn/start.effort and fails the turn.
Running at the default is a much better failure for config that is set once at
deploy time, where nobody is watching for it.

The accepted aliases are slackbotv2's. min, med, hi, xhi,
x-high all normalize the same way they do for -rsn, so a value that works in
a flag works in this config. Two surfaces disagreeing about the same word would
be worse than either one being stricter.

Scope

This is direction 1 for linearbot. Deliberately not included:

• githubbot. Same gap and the same fix shape, but its turn types are
  different (review, issue-work, CI-completed) and worth naming in their own
  PR rather than guessing at the taxonomy here.
• The Linear label mapping (direction 2). thinking-xhigh on an issue is
  the most useful half of linearbot/githubbot: per-turn-type reasoning-effort policy for autonomous turns (slackbotv2 has -rsn; no flag surface here) #1512, because it puts the decision where the
  knowledge is — a human triaging the ticket. It needs labels added to
  fetchLinearIssueContext's query plus a configurable prefix, and it composes
  cleanly on top of this: the label would override the per-turn-type default.
• -rsn in comment text (direction 3). Parity, and independent of both.

Happy to follow up with any of them; splitting keeps each reviewable.

Testing

Seven new tests: every codex effort accepted, slackbotv2's aliases accepted,
case and whitespace insensitivity, unrecognised and blank values dropped, the
two turn types resolving independently, one set not implying the other, and
unconfigured falling through to the harness default.

bun test in services/linearbot: 115 pass. bun run check:types clean.
helm lint passes; helm template renders each env only when its value is set.