route claim-swarm sub-agents by each issue's model:/effort: labels

[atomantic]

Summary

• Planners have been labeling issues model:/effort: and nothing consumed it. Every constant in dispatchLabels.js was producer-side — how to choose a label. The claim swarm fanned out N sub-agents and ran all of them at its own run's default, so model:heavy + effort:max bought the issue nothing.
• Adds the consumer half as two shared forms split by audience: DISPATCH_HINT_READING_GUIDANCE (what the labels mean — safe for any agent handed labeled issues) and DISPATCH_HINT_FANOUT_GUIDANCE (that plus the one line only an orchestrator can act on). The fan-out form is the reading form plus a line, never a second copy of the vocabulary, and both render their tiers/levels from DISPATCH_MODEL_TIERS / DISPATCH_EFFORT_LEVELS.
• Swarm block now carries each picked issue's labels out of Phase 1's listing through the partition step, tells the orchestrator to dispatch each fan-out agent at that issue's tier and budget, and requires the routing used to be named per issue in the run summary.
• Preloaded open-issues data input carries the reading form, outside the <portos-task-data> fence since it is PortOS instruction about how to read the block rather than forge data. Which inputs earn it is declared on the catalog entry (carriesDispatchLabels) rather than sniffed from the id; it is charged against MAX_TOTAL_CHARS; and it is skipped when the prompt already embeds it, so a swarm run with open-issues selected doesn't state the same contract twice.

Not in scope

branch-reconcile is the repo's second fan-out orchestrator and needs the same treatment, but it also needs a label fetch in formatInFlightForPrompt plus a PROMPT_VERSIONS bump — filed as #6373.

Test plan

• cd server && npm test — 2008 files, 39,914 passing.
• New coverage: the definition-site test pins both forms against the tier/level arrays and asserts the fan-out form is the reading form plus one line; the swarm test asserts the block carries the shared constant verbatim and the per-issue summary requirement; the data-input test asserts the paragraph appears only for issue-bearing inputs, sits outside the untrusted fence, is absent for a non-orchestrating consumer, and is not duplicated when the prompt already carries it.