Retire the planner and context pre-selection once the loop has delivered items

[thedancingdeveloper]

Part of #195. Blocked on #215, and deliberately not actionable until the loop has delivered items through the harness.

What becomes scaffolding

These exist only to prop up a model that gets one look:

• the planner — a whole model call per item whose sole job is to guess which files the implementer will need. A loop opens what it needs.
• context pre-selection (select_repo_context, ContextPolicy, --context-budget, --context-fallback-budget, The context budget is a target, not a ceiling: every item pays for the whole budget whatever it needs #152's ceiling work) — machinery to stop the planner over-filling a budget the loop does not have.
• edit blocks and to_diff (Name the text, do not count the lines (#195 Tier 1: parser and applier) #196, Ask the implementer for edit blocks, and compute the diff here (#195 Tier 1) #208, Tolerate a uniform indentation shift when it is still unique (#195 Tier 1) #210, Quote the file back when a SEARCH fails to match (#195 Tier 1) #211) and the patch apply ladder — a format for returning changes as text, when a loop leaves them in the tree.
• structured-text parsing and its format-error retries.

That is a large amount of code, most of it correct, all of it serving a shape that has never delivered an item.

Why this is not actionable yet

Deleting the only path that has tests, in favour of one that has never run in-harness, is the trade AGENTS.md rejects. The same staging was agreed for --session-host: land the replacement, prove it on a real backlog, then remove.

The trigger is evidence, not a date: items delivered through the role runner, through agent-harness run, with the gates and audit involved.

What must survive removal

• #152's insight is not scaffolding even though its mechanism is: a budget must be a ceiling, not a quota. A loop needs the same discipline expressed in turns and spend (Build the agentic role runner, and put the implementer through it #215).
• The checks and reviewer gates are untouched by any of this.
• _from_diff and everything after it stays: the loop's output still becomes a diff, a commit and a pull request.

An honest note on what is being deleted

Most of this was written or repaired on 2026-08-05, and several parts are genuinely good work — #216 found four real rendering defects in the edit-block path, including a CRLF file that could never be edited at all. Retiring it is not a judgement on the code. It is a judgement on the shape it serves.

Keeping it out of attachment would be the more expensive mistake: two execution paths mean two places for the gates to drift, which is the specific reason _from_diff has one body and two entrances.

Acceptance

• The planner role no longer exists, and nothing routes to it.
• run has one implement path.
• No test asserts behaviour that only the retired path had, unless the behaviour genuinely still matters.
• docs/ no longer describes context budgets as a thing an operator tunes per item.

Blind spots

• Whether any deployment still needs a single-shot path — an endpoint with no tool calling — is unknown. Claw Bay has tool calls; no other endpoint has been tried. If one turns up, this issue is wrong and Tier 1 becomes the floor it was originally argued to be.