feat(framework): collect business knowledge in the repo (#537)

[suleimansh]

Closes #537

Both bullets from the issue's Implementation section.

1. The docs go on the Context: line. .the-framework/README.md, DECISIONS.md and KNOWLEDGE-BASE.md join the #439 context line on every run, alongside the directories the user selects. One KNOWLEDGE_DOCS const; a test pins the post-merge prompt against it so the two halves cannot name different files.

2. The post-merge prompt asks the agent to update them. New ## Business knowledge section:

## Business knowledge

Consider whether the changes introduced by <SESSION_NAME> taught you something that belongs in these documents, and update them if so (create one if it doesn't exist yet):
- `.the-framework/README.md` (whole repo overview)
- `.the-framework/DECISIONS.md` (decisions taken, and why)
- `.the-framework/KNOWLEDGE-BASE.md` (business knowledge about the project)

Only write what a future agent would need and cannot get from the code itself.

BRAINSTORMING.md is left out, per the issue.

Two calls made here

• The docs ride with the built-in prompt, not with the user's dirs. Under --vanilla ("Disable system prompt") only the user's own --context dirs survive. The docs are ours, and [The Framework] Remove the persona / skill / memory framing #547 rule 3 is no framework-authored prompt when it's off.
• --eco-auto-maintenance no longer skips the whole post-merge run, it drops just ## Maintenance. Skipping wholesale was right while the post-merge prompt was exactly the maintenance section (Post-merge prompt: add TODO entries instead of running the presets inline #556). It isn't any more, and the flag does not name business knowledge, so it must not take it along.

Ran it for real

Real binary, real git repo, ~$1. A repo whose DECISIONS.md said "flat pricing, no surge" while the session's commit shipped surge pricing:

• The agent read DECISIONS.md off the context line unprompted, and caught the contradiction.
• It wrote a KNOWLEDGE-BASE.md that records what the code cannot: that the 1.4 multiplier has no recorded derivation, that bookedHour has no timezone attached, and that sameDay is caller-supplied so a client could opt out of surge. It explicitly left the surge rationale out, because that one is in a code comment.
• Both maintenance gates came back negative on a 7-line change, so no TODO file. The conditions work.

composeRunSystem 4475 -> 4573 chars.

Green: framework tests pass except runCli --daemon backgrounds the dashboard, which fails identically on clean main here (a daemon is already live on :4200) — pre-existing and environmental, not from this change.

[suleimansh]

Known limitation, tracked separately in #560: the post-merge run branches before writing, so the knowledge docs land on a branch nobody merges. Older than this PR (the #556 TODO entries do the same) and not widened into it.

[brillout]

• Let's remove README.md for now (very likely to be redundant with already existing README.md)
• Let's move the files to the repo root
• Add the comments also in the context

  - `DECISIONS.md` (decisions taken, and why)
  - `KNOWLEDGE-BASE.md` (knowledge and insights related to the project)

• post-merge is a misnommer, it should be called to something else e.g. on-before-mergeable
• I don't like the prompt, change it to:

  ## Business knowledge
  
  If you didn't already, consider updating the following based on the changes and discussions of ${{ tf.session_name }} (you can create the files if they're missing):
  - `DECISIONS.md` (decisions taken, and why)
  - `KNOWLEDGE-BASE.md` (knowledge and insights related to the project)
  
  Only write what a future agent would need and cannot get from the code itself.

@suleimansh I don't mind AI slope e.g. (DRY issues) for the code, but for all the prompting let's always take an extra look and give it extra thoughts. Thoughful prompts, always. Zero vibe-coded half-hassed prompts, ever. Feel free to ping me for everything related to prompting. When AI will apply the changes above, let's review carefuly whether its changes related to prompting matches what I said above (zero tolerence for AI slope in prompting).