Name FISHE_WEB_PORT when a web port is busy or misspelled

[dmccoystephenson]

Summary

• web/serve.py — the entry point the Dockerfile runs, and therefore the one most likely to be handed a value from outside — converted FISHE_WEB_PORT with a bare int(). A typo was reported as ValueError: invalid literal for int() with base 10: '80801x', which names neither the variable nor what it should hold. The integer check UserInterfaceFactory's WEB branch already had has been given to it as well.
• Neither web entry point handled a port that is simply taken. Both bound inside a constructor with nothing around it, so a second copy of the game — or anything else already on 8000/8080 — was surfaced as OSError: [Errno 98] Address already in use traced through http.server. allow_reuse_address does not cover this; only a socket left in TIME_WAIT is reopened by it, while a live listener still refuses the bind. A refused bind is now turned into a sentence naming the address, the reason (already listening / not permitted), and FISHE_WEB_PORT.
• Both failures happen before the front-end exists, so no showDialogue is available to say them through — they are raised, the way the factory already does.
• web/serve.py now binds before the URL is announced, so a failure is no longer preceded by an address that was never served.
• The two defaults are kept apart (8080 for the static server, 8000 for the server-backed front-end behind examples/web_app.py), since those are separate programs that can be run at the same time; a comment now records why. Only the check on the value is shared in spirit.
• README.md now mentions FISHE_WEB_PORT/FISHE_WEB_HOST, which the new messages instruct the player to set.

Front-ends touched

Only the two web entry points bind a socket, so this failure path has no console, pygame, or Pyodide equivalent: the console and pygame front-ends open no server, and the Pyodide front-end runs inside the browser tab. No BaseUserInterface primitive was changed, so front-end parity is unaffected.

Test plan

• python3 -m compileall -q src tests web
• python3 -m pytest — 820 passed, including four new tests in tests/web/test_serve.py (default port, port from the environment, a misspelled port naming the variable, a taken port explained rather than traced) and one in tests/ui/test_webUserInterface.py (a taken port explained rather than traced)
• black run over the changed files
• No Player/Stats/TimeService field changed, so schemas/*.json and the *JsonReaderWriter classes are untouched

Closes #176

This PR description was drafted during a Gardener session (https://github.com/Stephenson-Software/gardener).

[dmccoystephenson]

Self-review

The full diff was read and checked for logic errors, missed edge cases, front-end-parity gaps, schema/reader-writer mismatches, missing tests on public methods, and doc drift. No blocking issue was found. Three notes are recorded below, along with one gap that has already been closed on this branch.

Front-end parity — verified as not applicable rather than skipped. Only WebUserInterface and web/serve.py bind a socket: UserInterface/ConsoleUserInterface and PygameUserInterface open no server, and PyodideUserInterface runs inside the browser tab with no server at all. No BaseUserInterface primitive was added, removed, or changed, so nothing here has to be mirrored into the other three front-ends.

Save-file contract — no field on Player, Stats, or TimeService was touched, so schemas/player.json, schemas/stats.json, schemas/timeService.json and their *JsonReaderWriter classes are correctly left alone.

Socket cleanup on a failed bind — checked, because a helper that swallows and re-raises is where a half-open socket would be leaked. socketserver.TCPServer.__init__ already wraps server_bind()/server_activate() in a bare except that calls server_close() before re-raising, so the failed socket is closed before either _bindServer ever sees the OSError. Nothing further is needed.

Notes (non-blocking)

• src/ui/webUserInterface.py:165 — the raised OSError is constructed from a message alone, so errno and strerror are None on it. Nothing in the codebase catches this by errno, and raise ... from e keeps the original reachable via __cause__, so no behaviour is affected. The two-argument form (OSError(e.errno, message)) was considered and rejected: it would prefix the player-facing text with [Errno 98], which is the noise this change exists to remove. The same applies at web/serve.py:118.
• src/ui/webUserInterface.py:167 — the message tells the reader to set FISHE_WEB_PORT, while WebUserInterface also accepts port as a constructor argument that owes nothing to the environment (the tests pass it directly). In production the only construction path is UserInterfaceFactory's WEB branch, which does read that variable, so the advice is correct for every player who can see the message; it is only approximate for a caller constructing the class directly.
• web/serve.py:100 — the ValueError is re-raised from inside except ValueError without from None, so the underlying int() error is still shown as context. This is deliberate: UserInterfaceFactory's existing check at src/ui/userInterfaceFactory.py:46-50 does exactly the same, and the two are meant to behave identically.

Addressed during review

• tests/web/test_serve.py — _bindServer was covered only through its failure branch, so an implementation that always raised would have passed. A success-path test asserting that a free port is bound and the server handed back was added in the second commit on this branch.

This comment was drafted during a Gardener session (https://github.com/Stephenson-Software/gardener).