docs: write README links absolute so they resolve on PyPI

[lesnik512]

Implements modern-python/.github#85 for this repo.

Why

README.md is this package's PyPI long description, and PyPI does not rewrite relative links: it
resolves them against https://pypi.org/project/<name>/. Relative links here therefore 404 on the
package page while resolving correctly in the checkout, which is why the offline link gate is green
on them and always has been. The gate answers a different question than the package page asks.

Verified against a live rendered PyPI page rather than inferred from renderer behaviour. Measured
org-wide, 23 of 25 published packages carry relative README links, all dead on their package pages.

Design

Relative links become absolute blob/main/ self-links, or tree/main/ for a directory. The
--remap added in modern-python/.github#82 resolves those against the working tree, so the gate
keeps checking them rather than excluding them as external. Coverage is unchanged.

main rather than a release tag: a tag would stay correct in each frozen long description, but
modern-python/.github#82 deliberately excludes tag-pinned URLs from the gate, and writing the tag in
needs release-time substitution this repo does not have. main bounds the rot to old releases while
keeping every link checked.

The rule goes in AGENTS.md as an instruction rather than a test. CLAUDE.md imports AGENTS.md
here, so it is loaded into every agent turn, which makes it enforcement rather than documentation.
If a relative link appears in a README after this, that judgement is falsified and a test is the
answer. The <repo> placeholder keeps the paragraph byte-identical across all 25 repos.

Non-goals

• No test or lint check enforces the rule
• No change to relative links outside README.md, which are correct and gate-checked
• Not applied to .github or the two templates, which publish no long description

Verification

• No relative Markdown links remain in the README
• Offline link gate with this repo's remap: 0 errors
• Lint and tests are left to CI on this PR

[github-actions]

Benchmark gate

✅ gate passed

scenario	msg/s	delete/msg	WALrec/msg	WALB/msg	fpi	upd	del	dead_tup
consumer/w1/b10	797	1.000	6.81	909	242	5000	5000	10000
consumer/w1/b100	1101	1.000	6.85	917	243	5000	5000	10000
consumer/w2/b10	1382	1.000	6.79	958	242	5000	5000	10000
consumer/w2/b100	1401	1.000	6.74	954	243	5000	5000	10000
consumer/w4/b10	1027	1.000	6.79	980	261	5000	5000	10000
consumer/w4/b100	1937	1.000	6.85	1032	244	5000	5000	10000
consumer/w1/b100/tfbs100	7866	0.010	6.07	1116	243	5000	5000	10000
producer/w1/b100	3032	0.000	3.04	584	0	0	0	0

Gated (fails the build): delete_calls + tuple counters (upd/del/ins) + the producer's insert_calls, exact; select_calls within +2; wal_records within a 10% band. msg/s, WAL bytes and total calls are informational (timing/FPI noise).