Context window budget for a long-running task

I'm building something that processes {{describe volume of material}} to produce {{output}}. Model context: {{limit}}.

Design the context strategy:
1. Budget the window: what must always be present (instructions, schema, invariants), what's per-item, and what's retrieved. Give approximate token allocations that sum below the limit with headroom.
2. Decide the decomposition: map-reduce, sequential with running state, or retrieval-per-step. Justify against the failure mode of each — lost cross-references, drift in the running summary, or missed context in retrieval.
3. If state carries forward, define its exact structure and a hard size cap, plus what gets dropped when it's exceeded and by what rule.
4. Specify what must never be summarised because precision matters — numbers, IDs, quotes, definitions — and how those are carried verbatim.
5. Say where information loss will occur and what it costs the final output. Be explicit rather than optimistic.
6. Define the checkpointing: where results are persisted so a failure at item 400 doesn't restart everything.
7. Give the verification step that proves nothing was silently dropped — a count, a checksum, a coverage report.

Don't assume a bigger model solves this.